Appliid Math Text

Methods of Applied Mathematics
Lecture Notes
William G. Faris
May 14, 2002
2
Contents
1 Linear Algebra 7
1.1 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.1 Matrix algebra . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.2 Reduced row echelon form . . . . . . . . . . . . . . . . . . 8
1.1.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.1.4 The Jordan form . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.1.6 Quadratic forms . . . . . . . . . . . . . . . . . . . . . . . 14
1.1.7 Spectral theorem . . . . . . . . . . . . . . . . . . . . . . . 14
1.1.8 Circulant (convolution) matrices . . . . . . . . . . . . . . 15
1.1.9 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2 Vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.2.1 Vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.2.2 Linear transformations . . . . . . . . . . . . . . . . . . . . 18
1.2.3 Reduced row echelon form . . . . . . . . . . . . . . . . . . 18
1.2.4 Jordan form . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2.5 Forms and dual spaces . . . . . . . . . . . . . . . . . . . . 19
1.2.6 Quadratic forms . . . . . . . . . . . . . . . . . . . . . . . 20
1.2.7 Special relativity . . . . . . . . . . . . . . . . . . . . . . . 21
1.2.8 Scalar products and adjoint . . . . . . . . . . . . . . . . . 22
1.2.9 Spectral theorem . . . . . . . . . . . . . . . . . . . . . . . 23
1.3 Vector elds and dierential forms . . . . . . . . . . . . . . . . . 25
1.3.1 Coordinate systems . . . . . . . . . . . . . . . . . . . . . 25
1.3.2 Vector elds . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.3.3 Dierential forms . . . . . . . . . . . . . . . . . . . . . . . 27
1.3.4 Linearization of a vector eld near a zero . . . . . . . . . 28
1.3.5 Quadratic approximation to a function at a critical point 29
1.3.6 Dierential, gradient, and divergence . . . . . . . . . . . . 30
1.3.7 Spherical polar coordinates . . . . . . . . . . . . . . . . . 31
1.3.8 Gradient systems . . . . . . . . . . . . . . . . . . . . . . . 32
1.3.9 Hamiltonian systems . . . . . . . . . . . . . . . . . . . . . 33
1.3.10 Contravariant and covariant . . . . . . . . . . . . . . . . . 34
1.3.11 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3
4 CONTENTS
2 Fourier series 37
2.1 Orthonormal families . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2 L
2
convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Absolute convergence . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4 Pointwise convergence . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3 Fourier transforms 45
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 L
1
theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3 L
2
theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.4 Absolute convergence . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5 Fourier transform pairs . . . . . . . . . . . . . . . . . . . . . . . . 48
3.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.7 Poisson summation formula . . . . . . . . . . . . . . . . . . . . . 50
3.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.9 PDE Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4 Complex integration 53
4.1 Complex number quiz . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 Complex functions . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.2.1 Closed and exact forms . . . . . . . . . . . . . . . . . . . 54
4.2.2 Cauchy-Riemann equations . . . . . . . . . . . . . . . . . 55
4.2.3 The Cauchy integral theorem . . . . . . . . . . . . . . . . 55
4.2.4 Polar representation . . . . . . . . . . . . . . . . . . . . . 56
4.2.5 Branch cuts . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3 Complex integration and residue calculus . . . . . . . . . . . . . 57
4.3.1 The Cauchy integral formula . . . . . . . . . . . . . . . . 57
4.3.2 The residue calculus . . . . . . . . . . . . . . . . . . . . . 58
4.3.3 Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3.4 A residue calculation . . . . . . . . . . . . . . . . . . . . . 59
4.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.5 More residue calculus . . . . . . . . . . . . . . . . . . . . . . . . 60
4.5.1 Jordans lemma . . . . . . . . . . . . . . . . . . . . . . . . 60
4.5.2 A more delicate residue calculation . . . . . . . . . . . . . 61
4.5.3 Cauchy formula for derivatives . . . . . . . . . . . . . . . 61
4.5.4 Poles of higher order . . . . . . . . . . . . . . . . . . . . . 62
4.5.5 A residue calculation with a double pole . . . . . . . . . . 62
4.6 The Taylor expansion . . . . . . . . . . . . . . . . . . . . . . . . 63
4.6.1 Radius of convergence . . . . . . . . . . . . . . . . . . . . 63
4.6.2 Riemann surfaces . . . . . . . . . . . . . . . . . . . . . . . 64
CONTENTS 5
5 Distributions 67
5.1 Properties of distributions . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Mapping distributions . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 Radon measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Approximate delta functions . . . . . . . . . . . . . . . . . . . . . 70
5.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.6 Tempered distributions . . . . . . . . . . . . . . . . . . . . . . . . 72
5.7 Poisson equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.8 Diusion equation . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.9 Wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.10 Homogeneous solutions of the wave equation . . . . . . . . . . . 77
5.11 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.12 Answers to rst two problems . . . . . . . . . . . . . . . . . . . . 79
6 Bounded Operators 81
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2 Bounded linear operators . . . . . . . . . . . . . . . . . . . . . . 81
6.3 Compact operators . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.4 Hilbert-Schmidt operators . . . . . . . . . . . . . . . . . . . . . . 87
6.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.6 Finite rank operators . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7 Densely Dened Closed Operators 93
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2 Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.3 Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.4 Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.5 The spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.6 Spectra of inverse operators . . . . . . . . . . . . . . . . . . . . . 96
7.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.8 Self-adjoint operators . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.9 First order dierential operators with a bounded interval: point
spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.10 Spectral projection and reduced resolvent . . . . . . . . . . . . . 100
7.11 Generating second-order self-adjoint operators . . . . . . . . . . . 101
7.12 First order dierential operators with a semi-innite interval:
residual spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.13 First order dierential operators with an innite interval: contin-
uous spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.14 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.15 A pathological example . . . . . . . . . . . . . . . . . . . . . . . 104
6 CONTENTS
8 Normal operators 105
8.1 Spectrum of a normal operator . . . . . . . . . . . . . . . . . . . 105
8.2 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
8.3 Variation of parameters and Greens functions . . . . . . . . . . . 107
8.4 Second order dierential operators with a bounded interval: point
spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
8.5 Second order dierential operators with a semibounded interval:
continuous spectrum . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.6 Second order dierential operators with an innite interval: con-
tinuous spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
8.7 The spectral theorem for normal operators . . . . . . . . . . . . . 110
8.8 Examples: compact normal operators . . . . . . . . . . . . . . . 112
8.9 Examples: translation invariant operators and the Fourier trans-
form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
8.10 Examples: Schrodinger operators . . . . . . . . . . . . . . . . . . 113
8.11 Subnormal operators . . . . . . . . . . . . . . . . . . . . . . . . . 114
8.12 Examples: forward translation invariant operators and the Laplace
transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
8.13 Quantum mechanics . . . . . . . . . . . . . . . . . . . . . . . . . 118
8.14 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
9 Calculus of Variations 121
9.1 The Euler-Lagrange equation . . . . . . . . . . . . . . . . . . . . 121
9.2 A conservation law . . . . . . . . . . . . . . . . . . . . . . . . . . 122
9.3 Second variation . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
9.4 Interlude: The Legendre transform . . . . . . . . . . . . . . . . . 124
9.5 Lagrangian mechanics . . . . . . . . . . . . . . . . . . . . . . . . 126
9.6 Hamiltonian mechanics . . . . . . . . . . . . . . . . . . . . . . . . 127
9.7 Kinetic and potential energy . . . . . . . . . . . . . . . . . . . . . 128
9.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
9.9 The path integral . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
9.10 Appendix: Lagrange multipliers . . . . . . . . . . . . . . . . . . . 131
10 Perturbation theory 133
10.1 The implicit function theorem: scalar case . . . . . . . . . . . . . 133
10.2 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
10.3 The implicit function theorem: systems . . . . . . . . . . . . . . 136
10.4 Nonlinear dierential equations . . . . . . . . . . . . . . . . . . . 137
10.5 A singular perturbation example . . . . . . . . . . . . . . . . . . 138
10.6 Eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . . . . 139
10.7 The self-adjoint case . . . . . . . . . . . . . . . . . . . . . . . . . 141
10.8 The anharmonic oscillator . . . . . . . . . . . . . . . . . . . . . . 142
Chapter 1
Linear Algebra
1.1 Matrices
1.1.1 Matrix algebra
An m by n matrix A is an array of complex numbers A
ij
for 1 i m and
1 j n.
The vector space operations are the sum A+B and the scalar multiple cA.
Let A and B have the same dimensions. The operations are dened by
(A+B)
ij
= A
ij
+B
ij
(1.1)
and
(cA)
ij
= cA
ij
. (1.2)
The m by n zero matrix is dened by
0
ij
= 0. (1.3)
A matrix is a linear combination of other matrices if it is obtained from those
matrices by adding scalar multiples of those matrices.
Let A be an m by n matrix and B be an n by p matrix. Then the product
AB is dened by
(AB)
ik
=
n
j=1
A
ij
B
jk
. (1.4)
The n by n identity matrix is dened by
I
ij
=
ij
. (1.5)
Here
ij
is the Kronecker delta that is equal to 1 when i = j and to 0 when
i ,= j.
Matrix algebra satises the usual properties of addition and many of the
usual properties of multiplication. In particular, we have the associative law
(AB)C = A(BC) (1.6)
7
8 CHAPTER 1. LINEAR ALGEBRA
and the unit law
AI = IA = A. (1.7)
Even more important, we have the distributive laws
A(B +C) = AB +AC
(B +C)A = BA+CA. (1.8)
However multiplication is not commutative; in general AB ,= BA.
An n by n matrix A is invertible if there is another n by n matrix B with
AB = I and BA = I. In this case we write B = A
1
. (It turns out that if BA =
I, then also AB = I, but this is not obvious at rst.) The inverse operation has
several nice properties, including (A
1
)
1
= A and (AB)
1
= B
1
A
1
.
The notion of division is ambiguous. Suppose B is invertible. Then both
AB
1
and B
1
A exist, but they are usually not equal.
Let A be an n by n square matrix. The trace of A is the sum of the diagonal
entries. It is easy to check that tr(AB) = tr(BA) for all such matrices. Although
matrices do not commute, their traces do.
1.1.2 Reduced row echelon form
An m component vector is an m by 1 matrix. The ith standard basis vector is
the vector with 1 in the ith row and zeros everywhere else.
An m by n matrix R is in reduced row echelon form (rref) if each column
is either the next unit basis vector, or a a linear combination of the previous
unit basis vectors. The columns where unit basis vectors occur are called pivot
columns. The rank r of R is the number of pivot columns.
Theorem. If A is an m by n matrix, then there is an m by m matrix E that
is invertible and such that
EA = R, (1.9)
where R is in reduced row echelon form. The matrix R is uniquely determined
by A.
This theorem allows us to speak of the pivot columns of A and the rank of
A. Notice that if A is n by n and had rank n, then R is the identity matrix and
E is the inverse of A.
Example. Let
A =
_
_
4 12 2 16 1
0 0 3 12 2
1 3 0 2 3
_
_
. (1.10)
Then the rref of A is
R =
_
_
1 3 0 2 0
0 0 1 4 0
0 0 0 0 1
_
_
. (1.11)
Corollary. Let A have reduced row echelon form R. The null space of A is
the null space of R. That is, the solutions of the homogeneous equation Ax = 0
are the same as the solutions of the homogeneous equation Rx = 0.
1.1. MATRICES 9
Introduce a denition: A matrix is ipped rref if when ipped left-right
(iplr) and ipped up-down (ipud) it is rref.
The way reduced row echelon form is usually dened, one works from left
to right and from top to bottom. If you try to dene a corresponding concept
where you work from right to left and from bottom to top, a perhaps sensible
name for this is ipped reduced row echelon form.
Theorem. Let A be an m by n matrix with rank r. Then there is a unique
n by n r matrix N such that the transpose of N is ipped rref and such that
the transpose has pivot columns that are the non pivot columns of A and such
that
AN = 0 (1.12)
The columns of N are called the rational basis for the null space of A. It is
easy to nd N by solving RN = 0. The rational null space matrix N has the
property that its transpose is in ipped reduced row echelon form.
Example: In the above example the null space matrix of A is
N =
_
_
3 2
1 0
0 4
0 1
0 0
_
_
. (1.13)
That is, the solutions of Ax = 0 are the vectors of the form x = Nz. In other
words, the columns of N span the null space of A.
One can also use the technique to solve inhomogeneous equations Ax = b.
One simply applies the theory to the augmented matrix [A b]. There is a
solution when the last column of A is not a pivot column. A particular solution
may be read o from the last column of the rational basis.
Example. Solve Ax = b, where
b =
_
_
4
19
9
_
_
. (1.14)
To accomplish this, let
A
1
=
_
_
4 12 2 16 1 4
0 0 3 12 2 19
1 3 0 2 3 9
_
_
. (1.15)
Then the rref of A
1
is
R =
_
_
1 3 0 2 0 3
0 0 1 4 0 5
0 0 0 0 1 2
_
_
. (1.16)
The null space matrix of A
1
is
N
1
=
_
_
3 2 3
1 0 0
0 4 5
0 1 0
0 0 2
0 0 1
_
_
. (1.17)
Thus the solution of Ax = b is
x =
_
_
3
0
5
0
2
_
_
. (1.18)
1.1.3 Problems
Recall that the columns of a matrix A are linearly dependent if and only if the
homogeneous equation Ax = 0 has a non-trivial solution. Also, a vector y is in
the span of the columns if and only if the inhomogeneous equation Ax = y has
a solution.
1. Show that if p is a particular solution of the equation Ax = b, then every
other solution is of the form x = p +z, where Az = 0.
2. Consider the matrix
A =
_
_
2 4 5 19 3
2 4 3 9 2
4 8 1 17 2
3 6 1 4 2
4 8 1 7 3
2 4 3 21 0
_
_
.
Use Matlab to nd the reduced row echelon form of the matrix. Use this
to nd the rational basis for the solution of the homogeneous equation
Az = 0. Check this with the Matlab solution. Write the general solution
of the homogeneous equation.
3. Let A be the matrix given above. Use Matlab to nd an invertible matrix
E such that EA = R is in reduced echelon form. Find the determinant of
E.
4. Consider the system Ax = b, where A is as above, and
b =
_
_
36
21
9
6
6
23
_
_
.
1.1. MATRICES 11
Find the reduced echelon form of the matrix A augmented by the column
b on the right. Use this to nd the rational basis for the solution of the
homogeneous equation involving the augmented matrix. Use this to nd
the general solution of the original inhomogeneous equation.
5. Consider a system of 6 linear equations in 5 unknowns. It could be overde-
termined, that is, have no solutions. Or it could have special properties
and be under determined, that is, have many solutions. Could it be nei-
ther, that is, have exactly one solution? Is there such an example? Answer
this question, and prove that your answer is correct.
6. Consider a system of 5 linear equations in 6 unknowns. Could it have
exactly one solution? Is there such an example? Answer this question,
and prove that your answer is correct.
7. Consider the ve vectors in R
6
formed by the columns of A. Show that
the vector b is in the span of these ve vectors. Find explicit weights that
give it as a linear combination.
8. Is every vector y in R
6
in the span of these ve vectors? If so, prove it.
If not, give an explicit example of a vector that is not in the span, and
prove that it is not in the span.
9. Are these ve vectors linearly independent? Prove that your answer is
correct.
The vectors in A that are pivot columns of A have the same span as the
columns of A, yet are linearly independent. Find these vectors. How many
are they? Prove that they are linearly independent.
1.1.4 The Jordan form
Two matrices A, B are said to be similar if there is an invertible matrix P with
P
1
AP = B. Notice that if A and B are similar, then tr(A) = tr(B).
Theorem. Let A be an n by n matrix. Then there is an invertible matrix P
such that
P
1
AP = D +N, (1.19)
where D is diagonal with diagonal entries D
kk
=
k
. Each entry of N is zero,
except if
k
=
k1
, then it is allowed that N
k1 k
= 1.
Important note: Even if the matrix A is real, it may be that the matrices P
and D are complex.
The equation may also be written AP = PD + PN. If we let D
kk
=
k
,
then this is
n
j=1
A
ij
P
jk
= P
ik
k
+P
i k1
N
k1 k
. (1.20)
The N
k1 k
factor is zero, except in cases where
k
=
k1
, when it is allowed
to be 1. We can also write this in vector form. It is the eigenvalue equation
Au
k
=
k
u
k
(1.21)
except when
k
=
k1
, when it may take the form
Au
k
=
k
u
k
+u
k1
. (1.22)
The hard thing is to actually nd the eigenvalues
k
that form the diagonal
elements of the matrix D. This is a nonlinear problem. One way to see this is
the following. For each k = 1, . . . , n the matrix power A
k
is similar to a matrix
(D +N)
k
with the same diagonal entries as D
k
. Thus we have the identity
tr(A
k
) =
k
1
+ +
k
n
. (1.23)
This is a system of n nonlinear polynomial equations in n unknowns
1
, . . . ,
k
.
As is well known, there is another way of writing these equations in terms of the
characteristic polynomial. This gives a single nth order polynomial equation in
one unknown . This single equation has the same n solutions. For n = 2 the
equation is quite easy to deal with by hand. For larger n one is often driven to
computer algorithms.
Example: Let
A =
_
1 2
15 12
_
. (1.24)
The eigenvalues are 7, 6. The eigenvectors are the columns of
P =
_
1 2
3 5
_
. (1.25)
Let
D =
_
7 0
0 6
_
. (1.26)
Then AP = PD.
Example: Let
A =
_
10 1
9 4
_
. (1.27)
The only eigenvalue is 7. The eigenvector is the rst column of
P =
_
1 2
3 5
_
. (1.28)
Let
D +N =
_
7 1
0 7
_
. (1.29)
Then AP = P(D +N).
1.1. MATRICES 13
1.1.5 Problems
1. If A is a square matrix and f is a function dened by a convergent power
series, then f(A) is dened. Show that if A is similar to B, then f(A) is
similar to f(B).
2. By the Jordan form theorem, A is similar to D+N, where D is diagonal,
N is nilpotent, and D, N commute. To say that N is nilpotent is to say
that for some p 1 the power N
p
= 0. Show that
f(D +N) =
p1
m=0
1
m!
f
(m)
(D)N
m
(1.30)
3. Show that
exp(t(D +N)) = exp(tD)
p1
m=0
1
m!
N
m
t
m
. (1.31)
Use this to describe the set of all solutions x = exp(tA)z to the dierential
equation
dx
dt
= Ax. (1.32)
with initial condition x = z when t = 0.
4. Take
A =
_
0 1
k 2c
_
, (1.33)
where k 0 and c 0. The dierential equation describes an oscillator
with spring constant k and friction coecient 2c. Find the eigenvalues
and sketch a typical solution in the x
1
, x
2
plane in each of the following
cases: over damped c
2
> k > 0; critically damped c
2
= k > 0; under
damped 0 < c
2
< k; undamped 0 = c
2
< k; free motion 0 = c
2
= k.
5. Consider the critically damped case. Find the Jordan form of the matrix
A, and nd a similarity transformation that transforms A to its Jordan
form.
6. If A = PDP
1
, where the diagonal matrix D has diagonal entries
i
, then
f(A) may be dened for an arbitrary function f by f(A) = Pf(D)P
1
,
where f(D) is the diagonal matrix with entries f(
i
). Thus, for instance,
if each
i
0, then

A is dened. Find the square root of
A =
_
20 40
8 16
_
. (1.34)
7. Give an example of a matrix A with each eigenvalue
i
0, but for which
no square root

A can be dened? Why does the formula in the second
problem not work?
1.1.6 Quadratic forms
Given an m by n matrix A, its adjoint is an n by n matrix A
dened by
A
ij
=

A
ji
. (1.35)
If we are dealing with real numbers, then the adjoint is just the transpose. The
adjoint operator has several nice properties, including A
= A and (AB)
=
B
.
Two matrices A, B are said to be congruent if there is an invertible matrix
P with P
AP = B.
Theorem. Let A be an n by n matrix with A = A
. Then there is an
invertible matrix P such that
P
AP = D, (1.36)
where D is diagonal with entries 1, 1, and 0.
Dene the quadratic form Q(x) = x
Ax. Then the equation says that one

may make a change of variables x = Py so that Q(x) = y
Dy.
1.1.7 Spectral theorem
A matrix U is said to be unitary if U
1
= U
. In the real case this is the same

as being an orthogonal matrix.
Theorem. Let A be an n by n matrix with A = A
. Then there is a unitary

matrix U such that
U
1
AU = D, (1.37)
where D is diagonal with real entries.
Example: Let
A =
_
1 2
2 1
_
. (1.38)
The eigenvalues are given by
D =
_
3 0
0 1
_
. (1.39)
A suitable orthogonal matrix P is
P =
1
2
_
1 1
1 1
_
. (1.40)
Then AP = PD.
A matrix A is said to be normal if AA
= A
A. A self-adjoint matrix
(A
= A) is normal. A skew-adjoint matrix (A
= A) is normal. A unitary
matrix (A
= A
1
) is normal.
Theorem. Let A be an n by n matrix that is normal. Then there is a unitary
matrix U such that
U
1
AU = D, (1.41)
1.1. MATRICES 15
where D is diagonal.
Notice that this is the eigenvalue equation AU = UD. The columns of U
are the eigenvectors, and the diagonal entries of D are the eigenvalues.
Example: Let
P =
1
2
_
1 1
1 1
_
. (1.42)
Then P is a rotation by /4. The eigenvalues are on the diagonal of
F =
1
2
_
1 +i 0
0 1 i
_
. (1.43)
A suitable unitary matrix Q is
Q =
1
2
_
1 i
i 1
_
. (1.44)
Then PQ = QF.
1.1.8 Circulant (convolution) matrices
If A is normal, then there are unitary U and diagonal D with AU = UD. But
they are dicult to compute. Here is one special but important situation where
everything is explicit.
A circulant (convolution) matrix is an n by n matrix such that there is a
function a with A
pq
= a(p q), where the dierence is computed modulo n.
[For instance, a 4 by 4 matrix would have the same entry a(3) in the 12, 23, 34,
and 41 positions.]
The DFT (discrete Fourier transform) matrix is an n by n unitary matrix
given by
U
qk
=
1
n
e
2iqk
n
. (1.45)
Theorem. Let A be a circulant matrix. If U is the DFT matrix, then
U
1
AU = D, (1.46)
where D is a diagonal matrix with
D
kk
= a(k) =
r
a(r)e
2irk
n
. (1.47)
1.1.9 Problems
1. Let
A =
_
_
5 1 4 2 3
1 2 6 3 1
4 6 3 0 4
2 3 0 1 2
3 1 4 2 3
_
_
. (1.48)
Use Matlab to nd orthogonal P and diagonal D so that P
1
AP = D.
2. Find unitary Q and diagonal F so that Q
1
PQ = F.
3. The orthogonal matrix P is made of rotations in two planes and possibly
a reection. What are the two angles of rotations? Is there a reection
present as well? Explain.
4. Find an invertible matrix R such that R
AR = G, where G is diagonal
with all entries 1 or 0.
5. In the following problems the matrix A need not be square. Let A be a
matrix with trivial null space. Show that A
A is invertible.
6. Let E = A(A
A)
1
A
. Show that E
= E and E
2
= E.
7. Dene the pseudo-inverse A
+
= (A
A)
1
A
. Show that AA
+
= E and
A
+
A = I.
8. Let
A =
_
_
1 1 1
1 2 4
1 3 9
1 4 16
1 5 25
1 6 36
1 7 49
_
_
. (1.49)
Calculate E and check the identities above. Find the eigenvalues of E.
Explain the geometric meaning of E.
9. Find the parameter values x such that Ax best approximates
y =
_
_
7
5
6
1
3
7
24
_
_
. (1.50)
(Hint: Let x = A
+
y.) What is the geometric interpretation of Ax?
1.2. VECTOR SPACES 17
1.2 Vector spaces
1.2.1 Vector spaces
A vector space is a collection V of objects that may be combined with vector
addition and scalar multiplication. There will be two possibilities for the scalars.
They may be elements of the eld R of real numbers. Or they may be elements
of the eld C of complex numbers. To handle both cases together, we shall
consider the scalars as belonging to a eld F.
The vector space addition axioms say that the sum of each two vectors u, v
in V is another vector in the space called u+v in V . Addition must satisfy the
following axioms: associative law
(u +v) +w = u + (v +w) (1.51)
additive identity law
u +0 = 0 +u = u (1.52)
additive inverse law
u + (u) = (u) +u = 0 (1.53)
and commutative law
u +v = v +u. (1.54)
The vector space axioms also require that for each scalar c in F and each u
in V there is a vector cu in V . Scalar multiplication must satisfy the following
axioms: distributive law for vector addition
c(u +v) = cu +cv (1.55)
distributive law for scalar addition
(c +d)u = cu +cu (1.56)
associative law for scalar multiplication
(cd)u = c(du) (1.57)
identity law for scalar multiplication
1u = u. (1.58)
The elements of a vector space are pictured as arrows starting at a xed
origin. The vector sum is pictured in terms of the parallelogram law. The sum
of two arrows starting at the origin is the diagonal of the parallelogram formed
by the two arrows.
A subspace of a vector space is a subset that is itself a vector space. The
smallest possible subspace consists of the zero vector at the origin. A one
dimensional vector subspace is pictured as a line through the origin. A two
dimensional vector subspace is pictured as a plane through the origin, and so
on.
1.2.2 Linear transformations
Let V and W be vector spaces. A linear transformation T : V W is a function
from V to W that preserves the vector space operations. Thus
T(u +v) = Tu +Tv (1.59)
and
T(cu) = cTu. (1.60)
The null space of a linear transformation T : V W is the set of all vectors
u in V such that Tu = 0. The null space is sometimes called the kernel.
The range of a linear transformation T : V W is the set of all vectors Tu
in W, where u is in V . The range is sometimes called the image.
Theorem. The null space of a linear transformation T : V W is a subspace
of V .
Theorem. The range of a linear transformation T : V W is a subspace of
W.
Theorem. A linear transformation T : V W is one-to-one if and only if
its null space is the zero subspace.
Theorem. A linear transformation T : V W is onto W if and only if its
range is W.
Theorem. A linear transformation T : V W has an inverse T
1
: W V
if and only if its null space is the zero subspace and its range is W.
An invertible linear transformation will also be called a vector space isomor-
phism.
Consider a list of vectors p
1
, . . . , p
n
. They dene a linear transformation
P : F
n
V , by taking the vector in F
n
to be the weights of a linear combination
of the p
1
, . . . , p
n
.
Theorem. A list of vectors is linearly independent if and only if the corre-
sponding linear transformation has zero null space.
Theorem. A list of vectors spans the vector space V if and only if the
corresponding linear transformation has range V .
Theorem. A list of vectors is a basis for V if and only if the corresponding
linear transformation P : F
n
V is a vector space isomorphism. In that case
the inverse transformation P
1
: V F
n
is the transformation that carries a
vector to its coordinates with respect to the basis.
In case there is a basis transformation P : F
n
V , the vector space V is
said to be n dimensional.
1.2.3 Reduced row echelon form
Theorem Let p
1
, . . . , p
n
be a list of vectors in an m dimensional vector space
V . Let P : F
n
V be the corresponding transformation. Then there is an
isomorphism E : V F
m
such that
EP = R, (1.61)
where R is in reduced row echelon form. The m by n matrix R is uniquely
determined by the list of vectors.
Thus for an arbitrary list of vectors, there are certain vectors that are pivot
vectors. These vectors are those that are not linear combinations of previous
vectors in the list. The reduced row echelon form of a list of vectors expresses
the extent to which each vector in the list is a linear combination of previous
pivot vectors in the list.
1.2.4 Jordan form
Theorem. Let T : V V be a linear transformation of an n dimensional vector
space to itself. Then there is an invertible matrix P : R
n
V such that
P
1
TP = D +N, (1.62)
where D is diagonal with diagonal entries D
kk
=
k
. Each entry of N is zero,
except if
k
=
k1
, then it is allowed that N
k1 k
= 1.
We can also write this in vector form. The equation TP = P(D + N) is
equivalent to the eigenvalue equation
Tp
k
=
k
p
k
(1.63)
except when
k
=
k1
, when it may take the form
Tp
k
=
k
p
k
+p
k1
. (1.64)
It is dicult to picture such a linear transformation, even for a real vector
space. One way to do it works especially well in the case when the dimension of
V is 2. Then the linear transformation T maps each vector u in V to another
vector Tu in V . Think of each u as determining a point in the plane. Draw
the vector Tu as a vector starting at the point u. Then this gives a picture
of the linear transformation as a vector eld. The qualitative features of this
vector eld depend on whether the eigenvalues are real or occur as a complex
conjugate pair. In the real case, they can be both positive, both negative, or
of mixed sign. These give repelling, attracting, and hyperbolic xed points. In
the complex case the real part can be either positive or negative. In the rst
case the rotation has magnitude less than /2, and the vector eld is a repelling
spiral. In the second case the rotation has magnitude greater than /2 and the
vector eld is an attracting spiral. The case of complex eigenvalues with real
part equal to zero is special but important. The transformation is similar to a
rotation that has magnitude /2, and so the vector eld goes around in closed
elliptical paths.
1.2.5 Forms and dual spaces
The dual space V
of a vector space V consists of all linear transformations

from V to the eld of scalars F. If is in V
, then the value of on u in V is

written , u).
We think of the elements of F
n
as column vectors. The elements of the dual
space of F
n
are row vectors. The value of a row vector on a column vector is
given by taking the row vector on the left and the column vector on the right
and using matrix multiplication.
The way to picture an element of V
is via its contour lines. These are

straight lines at regularly spaced intervals. The value of on u is obtained by
taking the vector u and seeing how many contour lines it crosses.
Sometimes vectors in the original space V are called contravariant vectors,
while forms in the dual space V
are called covariant vectors.

1.2.6 Quadratic forms
A sesquilinear form is a function that assigns to each ordered pair of vectors u,
v in V a scalar a(u, v). It is required that it be linear in the second variable.
Thus
a(u, v +w) = a(u, v) +a(u, w) (1.65)
and
a(u, cv) = ca(u, v). (1.66)
Furthermore, it is required to be conjugate linear in the rst variable. this says
that
a(u +v, w) = a(u, w) +a(v, w) (1.67)
and
a(cu, v) = ca(u, v). (1.68)
In the real case one does not need the complex conjugates on the scalars. In
this case this is called a bilinear form.
A sesquilinear form denes a linear transformation from V to V
. (Actually,
in the complex case it is a conjugate linear transformation.) This linear trans-
formation takes the vector u to the form a(u, ). So a sesquilinear form may be
viewed as a special kind of linear transformation, but not one from a space to
itself.
A sesquilinear form is sometimes regarded as a covariant tensor. By contrast,
a linear transformation a mixed tensor.
A sesquilinear form is Hermitian if
a(u, v) = a(v, u). (1.69)
In the real case the complex conjugate does not occur. In this case the bilinear
form is called symmetric.
A Hermitian form denes a quadratic forma(u, u) that is real. The quadratic
form determines the Hermitian form. In fact,
4'a(u, v) = a(u +v, u +v) +a(u v, u v) (1.70)
determines the real part. Then a(u, v) = 'a(iu, v) determines the imaginary
part.
Theorem. Let a be a Hermitian form dened for vectors in an n dimensional
vector space V . Then there exists an invertible linear transformation P : F
n
V such that
a(Px, Px) = x
Dx, (1.71)
where D is diagonal with entries 1, 1, and 0.
It is sometimes possible to get a good picture of a quadratic form. In the
real case when V is two dimensional a quadratic form may be pictured by its
contour lines. There are dierent cases depending on whether one has +1s or
1s or a +1, 1 combination.
1.2.7 Special relativity
In special relativity V is the four dimensional real vector space of displacements
in space-time. Thus a vector u takes one event in space-time into another event
in space-time. It represents a separation between these space-time events.
There is a quadratic form g with signature +1, 1, 1, 1. The vector u is
said to be time-like, light-like, or space-like according to whether g(u, u) > 0,
g(u, u) = 0, or g(u, u) < 0. The time (perhaps in seconds) between time-like
separated events is
_
g(u, u). The distance (perhaps in light-seconds) between
space-like separated events is
_
g(u, u).
The light cone is the set of light-like vectors. The rest of the vector space
consists of three parts: forward time-like vectors, backward time-like vectors,
and space-like vectors.
Lemma. If two forward time-like vectors u, v have unit length
_
g(u, u) =
_
g(v, v) = 1, then their dierence u v is space-like.
Proof: This lemma is at least geometrically plausible. It may be proved by
using the diagonal form D of the quadratic form.
Lemma If u, v are forward time-like vectors with unit length, then g(u, v) >
1.
Proof: This follows immediately from the previous lemma.
Anti-Schwarz inequality. If u, v are forward time-like vectors, then g(u, v) >
_
g(u, u)
_
g(v, v).
Anti-triangle inequality. If u, v are forward time-like vectors, and u+v = w,
then
_
g(w, w) >
_
g(u, u) +
_
g(v, v). (1.72)
Proof: This follows immediately from the anti-Schwarz inequality.
The anti-triangle inequality is the basis of the twin paradox. In this case
the lengths of the vectors are measured in time units. One twin undergoes no
acceleration; the total displacement in space time is given by w. The other twin
begins a journey according to vector u, then changes course and returns to a
meeting event according to vector v. At this meeting the more adventurous
twin is younger.
The anti-triangle inequality is also the basis of the famous conversion of
mass to energy. In this case w is the energy-momentum vector of a particle,
and
_
g(w, w) is its rest mass. (Consider energy, momentum, and mass as all
measured in energy units.) The particle splits into two particles that y apart
from each other. The energy-momentum vectors of the two particles are u and
v. By conservation of energy-momentum, w = u + v. The total rest mass
_
g(u, u) +
_
g(v, v of the two particles is smaller than the original rest mass.
This has released the energy that allows the particles to y apart.
1.2.8 Scalar products and adjoint
A Hermitian form g(u, v) is an inner product (or scalar product) if g(u, u) > 0
except when u = 0. Often when there is just one scalar product in consideration
it is written g(u, v) = (u, v). However we shall continue to use the more explicit
notation for a while, since it is important to realize that to use these notions
there must be some mechanism that chooses an inner product. In the simplest
cases the inner product arises from elementary geometry, but there can be other
sources for a denition of a natural inner product.
Remark. An inner product gives a linear transformation from a vector space
to its dual space. Let V be a vector space with inner product g. Let u be in V .
Then there is a corresponding linear form u
in the dual space of V dened by

u
, v) = g(u, v) (1.73)
for all v in V . This form could also be denoted u
= gu.
The picture associated with this construction is that given a vector u repre-
sented by an arrow, the u
= gu is a form with contour lines perpendicular to

the arrow. The longer the vector, the more closely spaced the contour lines.
Theorem. (Riesz representation theorem) An inner product not only gives a
transformation from a vector space to its dual space, but this is an isomorphism.
Let V be a nite dimensional vector space with inner product g. Let : V F
be a linear form in the dual space of V . Then there is a unique vector
in V
such that
g(
, v) = , v) (1.74)
for all v in V . This vector could also be denoted
= g
1
.
The picture associated with this construction is that given a form repre-
sented by contour lines, the
= g
1
is a vector perpendicular to the lines.
The more closely spaced the contour lines, the longer the vector.
Theorem: Let V
1
and V
2
be two nite dimensional vector spaces with inner
products g
1
and g
2
. Let A : V
1
V
2
be a linear transformation. Then there is
another operator A
: V
2
V
1
such that
g
1
(A
u, v) = g
2
(Av, u) (1.75)
for every u in V
2
and v in V
1
. Equivalently,
g
1
(A
u, v) = g
2
(u, Av) (1.76)
for every u in V
2
and v in V
1
.
The outline of the proof of this theorem is the following. If u is given,
then the map g
2
(u, A) is an element of the dual space V
1
. It follows from the
previous theorem that it is represented by a vector. This vector is A
u.
A linear transformation A : V
1
V
2
is unitary if A
= A
1
.
There are more possibilities when the two spaces are the same. A linear
transformation A : V V is self-adjoint if A
= A. It is skew-adjoint if
A
= A. It is unitary if A
= A
1
.
A linear transformation A : V V is normal if AA
= A
A. Every operator
that is self-adjoint, skew-adjoint, or unitary is normal.
Remark 1. Here is a special case of the denition of adjoint. Fix a vector u
in V . Let A : F V be dened by Ac = cu. Then A
: V F is a form u
such that u
, z)c = g(z, cu). Thus justies the denition u
, z) = g(u, z).
Remark 2. Here is another special case of the denition of adjoint. Let
A : V F be given by a form . Then the adjoint A
: F V applied to
c is the vector c
such that g(c
, v) = c, v). This justies the denition

g(
, v) = , v).
There are standard equalities and inequalities for inner products. Two vec-
tors u, v are said to be orthogonal (or perpendicular) if g(u, v) = 0.
Theorem. (Pythagorus) If u, v are orthogonal and w = u +v, then
g(w, w) = g(u, u) +g(v, v). (1.77)
Lemma. If u, v are vectors with unit length, then 'g(u, v) 1.
Proof: Compute 0 g(u v, u v) = 2 2'g(u, v).
Lemma. If u, v are vectors with unit length, then [g(u, v)[ 1.
Proof: This follows easily from the previous lemma.
Notice that in the real case this says that 1 g(u, v) 1. This allows
us to dene the angle between two unit vectors u, v to be the number with
0 and cos() = g(u, v).
Schwarz inequality. We have the inequality [g(u, v)[
_
g(u, u)
_
g(v, v).
Triangle inequality. If u +v = w, then
_
g(w, w)
_
g(u, u) +
_
g(v, v). (1.78)
Proof: This follows immediately from the Schwarz inequality.
1.2.9 Spectral theorem
Theorem. Let V be an n dimensional vector space with an inner product g. Let
T : V V be a linear transformation that is normal. Then there is a unitary
transformation U : F
n
V such that
U
1
TU = D, (1.79)
where D is diagonal.
We can also write this in vector form. The transformation U denes an or-
thonormal basis of vectors. The equation is equivalent to the eigenvalue equation
TU = UD, or, explicitly,
Tu
k
=
k
u
k
. (1.80)
The distinction between linear transformations and quadratic forms some-
times shows up even in the context of matrix theory. Let A be a self-adjoint
matrix, regarded as dening a quadratic form. Let B be a self-adjoint matrix
that has all eigenvalues strictly positive. Then B denes a quadratic form that
is in fact a scalar product. The natural linear transformation to consider is then
T = B
1
A. The eigenvalue problem for this linear transformation then takes
the form
Au = Bu. (1.81)
Indeed, this version is often encountered in concrete problems. The linear trans-
formation B
1
A is a self-adjoint linear transformation relative to the quadratic
form dened by B. Therefore the eigenvalues are automatically real.
Example: Consider a system of linear spring oscillators. The equation of
motion is
M
d
2
x
dt
2
= Ax. (1.82)
Here M is the matrix of mass coecients (a diagonal matrix with strictly pos-
itive entries). The matrix A is real and symmetric, and it is assumed to have
positive eigenvalues. If we try a solution of the form x = z cos(
t), we get the

condition
Az = Mz. (1.83)
The should be real and positive, so the

indeed make sense as resonant
angular frequencies.
In this example the inner product is dened by the mass matrix M. The
matrix M
1
A is a self-adjoint operator relative to the inner product dened by
M. That is why the eigenvalues of M
1
A are real. In fact, the eigenvalues of the
linear transformation M
1
A are also positive. This is because the associated
quadratic form is dened by A, which is positive.
1.3. VECTOR FIELDS AND DIFFERENTIAL FORMS 25
1.3 Vector elds and dierential forms
1.3.1 Coordinate systems
In mathematical modeling it is important to be exible in the choice of coordi-
nate system. For this reason we shall consider what one can do with arbitrary
coordinate systems. We consider systems whose state can be described by two
coordinates. However the same ideas work for higher dimensional systems.
Say some part of the system can be described by coordinates u, v. Let ,
be another coordinate system for that part of the system. Then the scalars
and may be expressed as functions of the u and v. Furthermore, the matrix
J =
_

u
v
_
(1.84)
is invertible at each point, and the scalars u and v may be expressed as functions
of the and .
Consider the partial derivative /u as the derivative holding the coordinate
v constant. Similarly, the partial derivative /v is the derivative holding u
constant. Then the chain rule says that
u
=

u
+

u
. (1.85)
Similarly,
v
=

v
+

v
. (1.86)
These formulas are true when applied to an arbitrary scalar. Notice that these
formulas amount to the application of the transpose of the matrix J to the
partial derivatives.
One must be careful to note that an expression such as /u makes no
sense without reference to a coordinate system u, v. For instance, say that we
were interested in coordinate system u, . Then /u would be computed as a
derivative along a curve where is constant. This would be something quite
dierent. Thus we would have, for instance, by taking u = , the relation
u
=

u
+

u
. (1.87)
This seems like complete nonsense, until one realizes that the rst /u is taken
holding v constant, while the second /u is taken holding constant.
One way out of this kind of puzzle is to be very careful to specify the entire
coordinate system whenever a partial derivative is taken. Alternatively, one can
do as the chemists do and use notation that explicitly species what is being
held constant.
One can do similar computations with dierentials of the coordinates. Thus,
for instance,
du =
u
d +
u
d (1.88)
and
dv =
v
d +
v
d (1.89)
Dierentials such as these occur in line integrals. Notice that the relations are
quite dierent: they involve the inverse of the matrix J.
Here it is important to note that expressions such as du are dened quite
independently of the other members of the coordinate system. In fact, let h be
an arbitrary scalar. Then dh is dened and may be expressed in terms of an
arbitrary coordinate system. For instance
dh =
h
u
du +
h
v
dv (1.90)
and also
dh =
h
d +
h
d. (1.91)
1.3.2 Vector elds
For each point P in the space being studied, there is a vector space V (P),
called the tangent space at the point P. The elements of V (P) are pictured as
arrows starting at the point P. There are of course innitely many vectors in
each V (P). One can pick a linearly independent list that spans V (P) as basis
vectors for V (P).
We consider arbitrary coordinates x, y in the plane. These are not neces-
sarily Cartesian coordinates. Then there is a corresponding basis at each point
given by the partial derivatives /x and /y. This kind of basis is called a
coordinate basis.
A vector eld assigns to each point a vector in the tangent space at that
point. (The vector should depend on the point in a smooth way.) Thus a vector
eld is a rst order linear partial dierential operator of the form
L = a

x
+b

y
(1.92)
where a and b are functions of x and y. It is sometimes called a contravariant
vector eld.
With each vector eld L there is an associated ordinary dierential equation
dx
dt
= a = f(x, y) (1.93)
dy
dt
= b = g(x, y). (1.94)
Theorem. Let h be a scalar function of x and y. Then along each solution
of the dierential equation
dh
dt
= Lh. (1.95)
The vector eld determines the dierential equation. Furthermore, the dier-
ential equation determines the vector eld. They are the same thing, expressed
in somewhat dierent language. The vector eld or the dierential equation
may be expressed in an arbitrary coordinate system.
A solution of the dierential equation is a parameterized curve such that the
tangent velocity vector at each point of the curve is the value of the vector eld
at that point.
Theorem. (straightening out theorem) Consider a point where L ,= 0. Then
near that point there is a new coordinate system u and v such that
L =

u
. (1.96)
1.3.3 Dierential forms
A dierential form assigns to each point a linear form on the tangent space at
that point. (Again the assignment should be smooth.) Thus it is an expression
of the form
= p dx +q dy (1.97)
where p and q are functions of x and y. It is sometimes called a covariant vector
eld.
The value of the dierential form on the vector eld is the scalar function
, L) = pa +qb. (1.98)
It is a scalar function of x and y. Again, a dierential form may be expressed
in an arbitrary coordinate system.
A dierential form may be pictured in a rough way by indicating a bunch of
parallel lines near each point of the space. However, as we shall see, these lines
t together in a nice way only in a special situation (for an exact form).
A dierential form is said to be exact in a region if there is a smooth scalar
function h dened on the region such that
= dh. (1.99)
Explicitly, this says that
p dx +q dy =
h
x
dx +
h
y
dy. (1.100)
In this case the contour lines of the dierential form at each tangent space t
together on a small scale as pieces of the contour curves of the function h.
A dierential form is said to be closed in a region if it satises the integra-
bility condition
p
y
=
q
x
(1.101)
in the region.
Theorem. If a form is exact, then it is closed.
Theorem. (Poincare lemma) If a form is closed, then it is locally exact.
It is not true that if a form is closed in a region, then it is exact in the
region. Thus, for example, the form (y dx + xdy)/(x
2
+ y
2
) is closed in the
plane minus the origin, but it is not exact in this region. On the other hand,
it is locally exact, by the Poincare lemma. In fact, consider any smaller region,
obtained by removing a line from the origin to innity. In this smaller region
this form may be represented as d, where the angle has a discontinuity only
on the line.
When a dierential form is exact it is easy to picture. Then it is of the form
dh. The contour lines of h are curves. The dierential at any point is obtained
by zooming in at the point so close that the contour lines look straight.
A dierential form is what occurs as the integrand of a line integral. Let C
be an oriented curve. Then the line integral
_
C
=
_
C
p dx +q dy (1.102)
may be computed with an arbitrary coordinate system and with an arbitrary
parameterization of the oriented curve. The numerical answer is always the
same.
If a form is exact in a region, so that = dh, and if the curve C goes from
point P to point Q, then the integral
_
C
= h(Q) h(P). (1.103)
Thus the value of the line integral depends only on the end points. It turns out
that the form is exact in the region if and only if the value of the line integral
along each curve in the region depends only on the end points. Another way to
formulate this is to say that a form is exact in a region if and only if the line
integral around each closed curve is zero.
1.3.4 Linearization of a vector eld near a zero
The situation is more complicated at a point where L = 0. At such a point,
there is a linearization of the vector eld. Thus u, v are coordinates that vanish
at the point, and the linearization is given by
du
dt
=
a
x
u +
a
y
v (1.104)
dy
dt
=
b
x
u +
b
y
v. (1.105)
The partial derivatives are evaluated at the point. Thus this equation has the
matrix form
d
dt
_
u
v
_
=
_
a
x
a
y
b
x
b
y
_ _
u
v
_
. (1.106)
The idea is that the vector eld is somehow approximated by this linear vector
eld. The u and v represent deviations of the x and y from their values at this
point. This linear dierential equation can be analyzed using the Jordan form
of the matrix.
It is natural to ask whether at an isolated zero of the vector eld coordinates
can be chosen so that the linear equation is exactly equivalent to the original
non-linear equation. The answer is: usually, but not always. In particular,
there is a problem when the eigenvalues satisfy certain linear equations with
integer coecients, such as
1
+
2
= 0. This situation includes some of the
most interesting cases, such as that of conjugate pure imaginary eigenvalues.
The precise statement of the eigenvalue condition is in terms of integer linear
combinations m
1
1
+ m
2
2
with m
1
0, m
2
0, and m
1
+ m
2
2. The
condition is that neither
1
nor
2
may be expressed in this form.
How can this condition be violated? One possibility is to have
1
= 2
1
,
which says
1
= 0. Another is to have
1
= 2
2
. More interesting is
1
=
2
1
+
2
, which gives
1
+
2
= 0. This includes the case of conjugate pure
imaginary eigenvalues.
Theorem. (Sternberg linearization theorem). Suppose that a vector eld
satises the eigenvalue condition at a zero. Then there is a new coordinate
system near that zero in which the vector eld is a linear vector eld.
For the straightening out theorem and the Sternberg linearization theorem,
see E. Nelson, Topics in Dynamics I: Flows, Princeton University Press, Prince-
ton, NJ, 1969.
1.3.5 Quadratic approximation to a function at a critical
point
The useful way to picture a function h is by its contour lines.
Consider a function h such that at a particular point we have dh = 0. Then
there is a naturally dened quadratic form
d
2
h = h
xx
(dx)
2
+ 2h
xy
dxdy +h
yy
(dy)
2
. (1.107)
The partial derivatives are evaluated at the point. This quadratic form is called
the Hessian. It is determined by a symmetric matrix
H =
_
h
xx
h
xy
h
xy
h
yy
_
. (1.108)
Its value on a tangent vector L at the point is
d
2
h(L, L) = h
xx
a
2
+ 2h
xy
ab +h
yy
b
2
. (1.109)
Theorem (Morse lemma) Let h be such that dh vanishes at a point. Let h
0
be the value of h at this point. Suppose that the Hessian is non-degenerate at
the point. Then there is a coordinate system u, v such that near the point
h h
0
=
1
u
2
+
2
v
2
, (1.110)
where
1
and
2
are constants that are each equal to 1.
For the Morse lemma, see J. Milnor, Morse Theory, Princeton University
Press, Princeton, NJ, 1969.
1.3.6 Dierential, gradient, and divergence
It is important to distinguish between dierential and gradient. The dierential
of a scalar function h is given in an arbitrary coordinate system by
dh =
h
u
du +
h
v
dv. (1.111)
On the other hand, the denition of the gradient depends on the quadratic form
that denes the metric:
g = E(du)
2
+ 2F dudv +G(dv)
2
. (1.112)
It is obtained by applying the inverse g
1
of the quadratic form to the dierential
to obtain the corresponding vector eld. The inverse of the matrix
g =
_
E F
F G
_
(1.113)
is
g
1
=
1
[g[
_
G F
F E
_
, (1.114)
where [g[ = EGF
2
is the determinant of the matrix g. Thus
h =
1
[g[
_
G
h
u
F
h
v
_

u
+
1
[g[
_
F
h
u
+E
h
v
_

v
(1.115)
For orthogonal coordinates, when F = 0, the metric has the form g = E(du)
2
+
F(dv)
2
. In this case, the expression is simpler:
h =
1
E
_
h
u
_

u
+
1
G
_
h
v
_

v
(1.116)
Example: In polar coordinates
g = (dr)
2
+r
2
(d)
2
(1.117)
with determinant [g[ = r
2
. So
h =
h
r
r
+
1
r
2
h
. (1.118)
Another useful operation is the divergence of a vector eld. For a vector
eld
v = a

u
+b

v
(1.119)
the divergence is
v =
1
ga
u
+
1
gb
v
. (1.120)
In this formula
g denotes the square root of the determinant of the metric form

in this coordinate system. It is not immediately clear why this is the correct
quantity. However it may ultimately be traced back to the fact that the volume
element in an arbitrary coordinate system is

g dudv. For instance, in polar
coordinates it is r dr d.
Thus, for example, in polar coordinates the divergence of
v = a

r
+b

(1.121)
is
v =
1
r
ra
r
+
1
r
rb
=
1
r
ra
r
+
b
. (1.122)
The Laplace operator is the divergence of the gradient. Thus
h =
1
u
1
g
_
G
h
u
F
h
v
_
+
1
v
1
g
_
F
h
u
+E
h
v
_
. (1.123)
Again this simplies in orthogonal coordinates:
h =
1
g
E
h
u
+
1
g
G
h
v
. (1.124)
In polar coordinates this is
h =
1
r
r
_
r
h
r
_
+
1
r
2
2
h
2
. (1.125)
There are similar denitions in higher denitions. The important thing to
emphasize is that vector elds are very dierent from dierential forms, and this
dierence is highly visible once one leaves Cartesian coordinates. These objects
play dierent roles. A dierential form is the sort of object that ts naturally
into a line integral. On the other hand, a vector eld is something that denes
a ow, via the solution of a system of ordinary dierential equations.
1.3.7 Spherical polar coordinates
These ideas have analogs in higher dimensions. The dierential has the same
form in any coordinate system. On the other hand, the gradient depends on the
inner product.
Example: Consider spherical polar coordinates, where is co-latitude and
is longitude. Then the metric form is
g = (dr)
2
+r
2
(d)
2
+r
2
sin
2
()(d)
2
. (1.126)
In this case the gradient is
h =
h
r
r
+
1
r
2
h
+
1
r
2
sin
2
()
h
. (1.127)
The divergence of a vector eld makes sense in any number of dimensions.
For spherical polar coordinates the volume element is r
2
sin() dr d d. The
divergence of
v = a

r
+b

+c

(1.128)
is
v =
1
r
2
sin()
r
2
sin()a
r
+
1
r
2
sin()
r
2
sin()b
+
1
r
2
sin()
r
2
sin()c
.
(1.129)
This can be written more simply as
v =
1
r
2
r
2
a
r
+
1
sin()
sin()b
+
c
. (1.130)
The Laplace operator is the divergence of the gradient. In spherical polar
coordinates this is
h =
1
r
2
r
_
r
2
h
r
_
+
1
r
2
sin()
sin()
h
+
1
r
2
sin
2
()
2
h
2
. (1.131)
1.3.8 Gradient systems
Now we use Cartesian coordinates x, y and the usual inner product. This inner
product maps a form to a corresponding vector eld. Thus the dierential form
dh =
h
x
dx +
h
y
dy (1.132)
determines a gradient vector eld
h =
h
x
x
+
h
y
y
. (1.133)
The corresponding dierential equation is
dx
dt
= h
x
(1.134)
dy
dt
= h
y
. (1.135)
Theorem. Along a solution of a gradient system
dh
dt
= h
2
x
+h
2
y
0. (1.136)
This is hill climbing. The picture is that at each point, the gradient vector
eld points uphill in the steepest direction.
Consider a point where dh = 0. Then the linearization of the equation is
given by
du
dt
= h
xx
u +h
xy
v (1.137)
dv
dt
= h
xy
u +h
yy
v. (1.138)
where the partial derivatives are evaluated at the point. This is a real symmetric
matrix, and therefore its eigenvalues are real.
1.3.9 Hamiltonian systems
Now we use Cartesian coordinates x, y, but now instead of a symmetric inner
product we use a skew-symmetric form to map a form to a corresponding vector.
Thus the dierential form
dH =
H
x
dx +
H
y
dy (1.139)
determines a Hamiltonian vector eld
H =
H
y
x

H
x
y
. (1.140)
The corresponding dierential equation is
dx
dt
= H
y
(1.141)
dy
dt
= H
x
. (1.142)
Theorem. (Conservation of Energy) Along a solution of a Hamiltonian sys-
tem.
dH
dt
= H
y
H
x
H
x
H
y
= 0. (1.143)
This is conservation of energy. The picture is that at each point the vector
eld points along one of the contour lines of H.
Consider a point where dH = 0. Then the linearization of the equation is
given by
du
dt
= H
xy
u +H
yy
v (1.144)
dv
dt
= H
xx
u H
xy
v. (1.145)
where the partial derivatives are evaluated at the point. This is a special kind
of matrix. The eigenvalues satisfy the equation
2
= H
2
xy
H
xx
H
yy
. (1.146)
So they are either real with opposite sign, or they are complex conjugate pure
imaginary. In the real case solutions linger near the equilibrium point. In the
complex case solutions oscillate near the equilibrium point.
1.3.10 Contravariant and covariant
Here is a summary that may be useful. Contravariant objects include points
and vectors (arrows). These are objects that live in the space. They also include
oriented curves (whose tangent vectors are arrows). Thus a vector eld (con-
travariant) determines solution curves of the associated system of dierential
equations (contravariant).
Covariant objects associate numbers to objects that live in the space. They
include scalars, linear forms (elements of the dual space), and bilinear forms.
They also include scalar functions and dierential forms. Often covariant objects
are described by their contours. The dierential of a scalar function (covariant)
is a dierential form (covariant). The integral of a dierential form (a covariant
object) along an oriented curve (a contravariant object) is a number.
There are also mixed objects. A linear transformation is a mixed object,
since it takes vectors as inputs (covariant), but it gives vectors as outputs (con-
travariant).
A bilinear form may be used to move from the contravariant world to the
covariant world. The input is a vector, and the output is a linear form. If the
bilinear form is non-degenerate, then there is an inverse that can take a linear
form as an input and give a vector as an output. This occurs for instance when
we take the dierential of a scalar function (covariant) and use a symmetric
bilinear form to get a gradient vector eld (contravariant). It also occurs when
we take the dierential of a scalar function (covariant) and use an anti-symmetric
bilinear form to get a Hamiltonian vector eld (contravariant).
1.3.11 Problems
1. A vector eld that is non-zero at a point can be transformed into a constant
vector eld near that point. Pick a point away from the origin and nd
coordinates u and v near there so that
y

x
+x

y
=

u
. (1.147)
2. A vector eld that is non-zero at a point can be transformed into a constant
vector eld near that point. Pick a point away from the origin and nd
coordinates u and v so that
y
x
2
+y
2
x
+
x
x
2
+y
2
y
=

u
. (1.148)
3. A dierential form usually cannot be transformed into a constant dier-
ential form, but there special circumstances when that can occur. Is it
possible to nd coordinates u and v near a given point (not the origin) so
that
y dx +xdy = du? (1.149)
4. A dierential form usually cannot be transformed into a constant dier-
ential form, but there special circumstances when that can occur. Is it
possible to nd coordinates u and v near a given point (not the origin) so
that
y
x
2
+y
2
dx +
x
x
2
+y
2
dy = du? (1.150)
5. Consider the vector eld
L = x(4 x y)

x
+ (x 2)y

y
. (1.151)
Find its zeros. At each zero, nd the linearization. For each linearization,
nd the eigenvalues. Use this information to sketch the vector eld.
6. Let h be a smooth function. Its gradient expressed with respect to Carte-
sian basis vectors /x and /y is
h =
h
x
x
+
h
y
y
. (1.152)
Find the gradient h expressed with respect to polar basis vectors /r
and /.
7. Let H be a smooth function. Its Hamiltonian vector eld expressed with
respect to Cartesian basis vectors /x and /y is
H =
H
y
x

H
x
y
. (1.153)
Find this same vector eld expressed with respect to polar basis vectors
/r and /.
8. Consider the vector eld
L = (1 +x
2
+y
2
)y

x
(1 +x
2
+y
2
)x

y
. (1.154)
Find its linearization at 0. Show that there is no coordinate system near
0 in which the vector eld L is expressed by its linearization. Hint: Solve
the associated system of ordinary dierential equations, both for L and
for its linearization. Find the period of a solution in both cases.
9. Here is an example of a xed point where the linear stability analysis
gives an elliptic xed point but changing to polar coordinates shows the
unstable nature of the xed point:
dx
dt
= y +x(x
2
+y
2
) (1.155)
dy
dt
= x +y(x
2
+y
2
). (1.156)
Change the vector eld to the polar coordinate representation and solve
the corresponding system of ordinary dierential equations.
Chapter 2
Fourier series
2.1 Orthonormal families
Let T be the circle parameterized by [0, 2) or by [, ). Let f be a complex
function on T that is integrable. The nth Fourier coecient is
c
n
=
1
2
_
2
0
e
inx
f(x) dx. (2.1)
The goal is to show that f has a representation as a Fourier series
f(x) =
n=
c
n
e
inx
. (2.2)
There are two problems. One is to interpret the sense in which the series
converges. The second is to show that it actually converges to f.
This is a huge subject. However the simplest and most useful theory is in
the context of Hilbert space. Let L
2
(T) be the space of all (Borel measurable)
functions such that
|f|
2
=
1
2
_
2
0
[f(x)[
2
dx < . (2.3)
Then L
2
(T) is a Hilbert space with inner product
(f, g) =
1
2
_
2
0
f(x)g(x) dx. (2.4)
Let
n
(x) = exp(inx). (2.5)
Then the
n
form an orthogonal family in L
2
(T). Furthermore, we have the
identity
|f|
2
=
|n|N
[c
n
[
2
+|f
|n|N
c
n
n
|
2
. (2.6)
37
38 CHAPTER 2. FOURIER SERIES
In particular, we have Bessels inequality
|f|
2
|n|N
[c
n
[
2
. (2.7)
This shows that
n=
[c
n
[
2
< . (2.8)
The space of sequences satisfying this identity is called
2
. Thus we have proved
the following theorem.
Theorem. If f is in L
2
(T), then its sequence of Fourier coecients is in
2
.
2.2 L
2
convergence
It is easy to see that

|n|N
c
n
n
is a Cauchy sequence in L
2
(T) as N .
It follows from the completeness of L
2
(T) that it converges in the L
2
sense to a
sum g in L
2
(T). That is, we have
lim
N
|g
|n|N
c
n
n
|
2
= 0. (2.9)
The only remaining thing to show is that g = f. This, however, requires some
additional ideas.
One way to do it is to take r with 0 < r < 1 and look at the sum
n=
r
|n|
c
n
e
inx
=
1
2
_
2
0
P
r
(x y)f(y) dy. (2.10)
Here
P
r
(x) =
n=
r
|n|
e
inx
=
1 r
2
1 2r cos(x) +r
2
. (2.11)
The functions
1
2
P
r
(x) have the properties of an approximate delta function.
Each such function is positive and has integral 1 over the periodic interval.
Furthermore,
P
r
(x)
1 r
2
2r(1 cos(x))
, (2.12)
which approaches zero as r 1 away from points where cos(x) = 1.
Theorem. If f is continuous on T, then
f(x) = lim
r1
n=
r
|n|
c
n
e
inx
, (2.13)
and the convergence is uniform.
2.2. L
2
CONVERGENCE 39
Proof: It is easy to compute that
n=
r
|n|
c
n
e
inx
=
1
2
_
2
0
P
r
(xy)f(y) dy =
1
2
_
2
0
P
r
(z)f(xz) dz. (2.14)
Hence
f(x)
n=
r
|n|
c
n
e
inx
=
1
2
_
2
0
P
r
(z)(f(x) f(x z)) dz. (2.15)
Thus
[f(x)
n=
r
|n|
c
n
e
inx
[ = [
1
2
_
2
0
P
r
(z)(f(x)f(xz)) dz[
1
2
_
2
0
P
r
(z)[f(x)f(xz)[ dz.
(2.16)
Since f is continuous on T, its absolute value is bounded by some constant
M. Furthermore, since f is continuous on T, it is uniformly continuous on T.
Consider arbitrary > 0. It follows from the uniform continuity that there exists
> 0 such that for all x and all z with [z[ < we have [f(x) f(x z)[ < /2.
Thus the right hand side is bounded by
1
2
_
|z|<
P
r
(z)[f(x)f(xz)[ dz+
1
2
_
|z|
P
r
(z)[f(x)f(xz)[ dx /2+
1 r
2
2r(1 cos())
2M.
(2.17)
Take r so close to 1 that the second term is also less than /2. Then the dierence
that is being estimated is less than for such r. This proves the result.
2
(T), then
f =
n=
c
n
n
(2.18)
in the sense that
lim
N
|f
|n|N
c
n
n
|
2
= 0. (2.19)
Proof: Let f have Fourier coecients c
n
and g =

n
c
n
n
. Let > 0 be
arbitrary. Let h be in C(T) with
|f h|
2
< /2. (2.20)
Suppose h has Fourier coecients d
n
. Then there exists r < 1 such that
|h
n
r
n
d
n
e
inx
|
< /2. (2.21)

It follows that
|h
n
r
n
d
n
e
inx
|
2
< /2. (2.22)
Thus
|f
n
r
n
d
n
e
inx
|
2
< . (2.23)
However since g is the orthogonal projection of f onto the span of the
n
, it is
also the best approximation of f in the span of the
n
. Thus
|f g|
2
< . (2.24)
Since > 0 is arbitrary, we can only conclude that f = g.
2
(T), then
|f|
2
=
n=
[c
n
[
2
. (2.25)
2.3 Absolute convergence
Dene the function spaces
C(T) L
(T) L
2
(T) L
1
(T). (2.26)
The norms |f|
on the rst two spaces are the same, the smallest number M
such that [f(x)[ M (with the possible exception of a set of x of measure zero).
The space C(T) consists of continuous functions; the space L
(T) consists of
all bounded functions. The norm on L
2
(T) is given by |f|
2
2
=
1
2
_
2
0
[f(x)[
2
dx.
The norm on L
1
(T) is given by |f|
1
=
1
2
_
2
0
[f(x)[ dx. Since the integral is a
probability average, their relation is
|f|
1
|f|
2
|f|
. (2.27)
Also dene the sequence spaces
1

2
c
0

. (2.28)
The norm on
1
is |c|
1
=

n
[c
n
[. Then norm on
2
is given by |c|
2
2
=

n
[c
n
[
2
.
The norms on the last two spaces are the same, that is, |c|
is the smallest
M such that [c
n
[ M. The space c
0
consists of all sequences with limit 0 at
innity. The relation between these norms is
|c|
|c|
2
|c|
1
. (2.29)
We have seen that the Fourier series theorem gives a perfect correspondence
between L
2
(T) and
2
. For the other spaces the situation is more complex.
Riemann-Lebesgue Lemma. If f is in L
1
(T), then the Fourier coecients of
f are in c
0
, that is, they approaches 0 at innity.
Proof: Each function in L
2
(T) has Fourier coecients in
2
, so each function
in L
2
(T) has Fourier coecients that vanish at innity. The map from a function
to its Fourier coecients gives a continuous map from L
1
(T) to
. However
2.4. POINTWISE CONVERGENCE 41
every function in L
1
(T) may be approximated arbitrarily closely in L
1
(T) norm
by a function in L
2
(T). Hence its coecients may be approximated arbitrarily
well in
norm by coecients that vanish at innity. Therefore the coecients

vanish at innity.
In summary, the map from a function to its Fourier coecients gives a con-
tinuous map from L
1
(T) to c
0
. That is, the Fourier coecients of an integrable
function are bounded (this is obvious) and approach zero (Riemann-Lebesgue
lemma). Furthermore, the Fourier coecients determine the function.
The map from Fourier coecients to functions gives a continuous map from
1
to C(T). An sequence that is absolutely summable denes a Fourier series
that converges absolutely and uniformly to a continuous function.
2
(T) and if f
exists (in the sense that f is an integral

of f) and if f
is also in L
2
(T), then the Fourier coecients are in
1
.
2.4 Pointwise convergence
This gives a very satisfactory picture of Fourier series. However there is one
slightly unsatisfying point. The convergence in the L
2
sense does not imply
convergence at a particular point. Of course, if the derivative is in L
2
then
we have uniform convergence, and in particular convergence at each point. But
what if the function is dierentiable at one point but has discontinuities at other
points? What can we say about convergence at that one point? Fortunately, we
can nd something about that case by a closer examination of the partial sums.
One looks at the partial sum
|n|N
c
n
e
inx
=
1
2
_
2
0
D
N
(x y)f(y) dy. (2.30)
Here
D
N
(x) =
|n|N
e
inx
=
sin((N +
1
2
)x)
sin(
1
2
x)
. (2.31)
This Dirichlet kernel D
N
(x) has at least some of the properties of an approxi-
mate delta function. Unfortunately, it is not positive; instead it oscillates wildly
for large N at points away from where sin(x/2) = 0. However the function
1/(2)D
N
(x) does have integral 1.
Theorem. If for some x the function
d
x
(z) =
f(x +z) f(x)
2 sin(z/2)
(2.32)
is in L
1
(T), then at that point
f(x) =
n=
c
n
n
(x). (2.33)
Note that if d
x
(z) is continuous at z = 0, then its value at z = 0 is d
x
(0) =
f
(x). So the hypothesis of the theorem is a condition related to dierentiability

of f at the point x. The conclusion of the theorem is pointwise convergence of
the Fourier series at that point. Since f may be discontinuous at other points,
it is possible that this Fourier series is not absolutely convergent.
Proof: We have
f(x)
|n|N
c
n
e
inx
=
1
2
_
2
0
D
N
(z)(f(x) f(x z)) dz. (2.34)
We can write this as
f(x)
|n|N
c
n
e
inx
=
1
2
_
2
0
2 sin((N +
1
2
)z)d
x
(z) dz. (2.35)
This goes to zero as N , by the Riemann-Lebesgue lemma.
2.5 Problems
1. Let f(x) = x dened for x < . Find the L
1
(T), L
2
(T), and L
(T)
norms of f, and compare them.
2. Find the Fourier coecients c
n
of f for all n in Z.
3. Find the
,
2
, and
1
norms of these Fourier coecients, and compare
them.
4. Use the equality of L
2
and
2
norms to compute
(2) =
n=1
1
n
2
.
5. Compare the
and L
1
norms for this problem. Compare the L
and
1
norms for this problem.
6. Use the pointwise convergence at x = /2 to evaluate the innite sum
k=0
(1)
k
1
2k + 1
,
regarded as a limit of partial sums. Does this sum converge absolutely?
7. Let F(x) =
1
2
x
2
dened for x < . Find the Fourier coecients of
this function.
8. Use the equality of L
2
and
2
norms to compute
(4) =
n=1
1
n
4
.
2.5. PROBLEMS 43
9. Compare the
and L
1
norms for this problem. Compare the L
and
1
norms for this problem.
10. At which points x of T is F(x) continuous? Dierentiable? At which
points x of T is f(x) continuous? Dierentiable? At which x does F
(x) =
f(x)? Can the Fourier series of f(x) be obtained by dierentiating the
Fourier series of F(x) pointwise? (This last question can be answered by
inspecting the explicit form of the Fourier series for the two problems.)
Chapter 3
Fourier transforms
3.1 Introduction
Let R be the line parameterized by x. Let f be a complex function on R that
is integrable. The Fourier transform

f = Ff is
f(k) =
_

e
ikx
f(x) dx. (3.1)
It is a function on the (dual) real line R
parameterized by k. The goal is to

show that f has a representation as an inverse Fourier transform
f(x) =
_

e
ikx

f(k)
dk
2
. (3.2)
There are two problems. One is to interpret the sense in which these integrals
converge. The second is to show that the inversion formula actually holds.
The simplest and most useful theory is in the context of Hilbert space. Let
L
2
(R) be the space of all (Borel measurable) complex functions such that
|f|
2
2
=
_

[f(x)[
2
dx < . (3.3)
Then L
2
(R) is a Hilbert space with inner product
(f, g) =
_
f(x)g(x) dx. (3.4)
Let L
2
(R
) be the space of all (Borel measurable) complex functions such that

|h|
2
2
=
_

[h(k))[
2
dk
2
< . (3.5)
Then L
2
(R
) is a Hilbert space with inner product

(h, u) =
_

h(k)u(k)
dk
2
. (3.6)
45
46 CHAPTER 3. FOURIER TRANSFORMS
We shall see that the correspondence between f and

f is a unitary map from
L
2
(R) onto L
2
(R
). So this theory is simple and powerful.

3.2 L
1
theory
First, we need to develop the L
1
theory. The space L
1
is a Banach space. Its
dual space is L
, the space of essentially bounded functions. An example of a

function in the dual space is the exponential function
k
(x) = e
ikx
. The Fourier
transform is then
f(k) =
k
, f) =
_

k
(x)f(x) dx, (3.7)
where
k
is in L
and f is in L
1
.
Theorem. If f, g are in L
1
(R), then the convolution f g is another function
in L
1
(R) dened by
(f g)(x) =
_

f(x y)g(y) dy. (3.8)

Theorem. If f, g are in L
1
(R), then the Fourier transform of the convolution
is the product of the Fourier transforms:
(f g)(k) =

f(k) g(k). (3.9)
Theorem. Let f
(x) = f(x). Then the Fourier transform of f
is the
complex conjugate of

f.
1
(R), then its Fourier transform

f is in L
(R
) and
satises |
f|
|f|
1
. Furthermore,

f is in C
0
(R
), the space of bounded

continuous functions that vanish at innity.
1
and is also continuous and bounded, we have the
inversion formula in the form
f(x) = lim
0
_

e
ikx
(k)

f(k)
dk
2
, (3.10)
where
(k) = exp([k[). (3.11)

Proof: The inverse Fourier transform of this is
(x) =
1
x
2
+
2
. (3.12)
It is easy to calculate that
_

e
ikx
(k)

f(k)
dk
2
= (
f)(x). (3.13)
However
is an approximate delta function. The result follows by taking 0.

3.3. L
2
THEORY 47
3.3 L
2
theory
The space L
2
is its own dual space, and it is a Hilbert space. It is the setting
for the most elegant and simple theory of the Fourier transform.
Lemma. If f is in L
1
(R) and in L
2
(R), then

f is in L
2
(R
), and |f|
2
2
= |
f|
2
2
.
Proof. Let g = f
f. Then g is in L
1
and is continuous and bounded.
Furthermore, the Fourier transform of g is [
f(k)[
2
. Thus
|f|
2
2
= g(0) = lim
0
_

(k)[
f(k)[
2
dk
2
=
_

f(k)[
2
dk
2
. (3.14)
Theorem. Let f be in L
2
(R). For each a, let f
a
= 1
[a,a]
f. Then f
a
is in
L
1
(R) and in L
2
(R), and f
a
f in L
2
(R) as a . Furthermore, there exists
f in L
2
(R
) such that

f
a

f as a .
Explicitly, this says that the Fourier transform

f(k) is characterized by
_

f(k)
_
a
a
e
ikx
f(x) dx[
2
dk
2
0 (3.15)
as a .
These arguments show that the Fourier transformation F : L
2
(R) L
2
(R
)
dened by Ff =

f is well-dened and preserves norm. It is easy to see from
the fact that it preserves norm that it also preserves inner product: (Ff, Fg) =
(f, g).
Dene the inverse Fourier transform F
in the same way, so that if h is in

L
1
(R
) and in L
2
(R
), then F
h is in L
2
(R) and is given by the usual inverse
Fourier transform formula. Again we can extend the inverse transformation to
F
: L
2
(R
) L
2
(R) so that it preserves norm and inner product.
Now it is easy to check that (F
h, f) = (h, Ff). Take h = Fg. Then

(F
Fg, f) = (Fg, Ff) = (g, f). That is F
Fg = g. Similarly, one may show

that FF
u = u. These equations show that F is unitary and that F
= F
1
is
the inverse of F. This proves the following result.
Theorem. The Fourier transform F initially dened on L
1
(R) L
2
(R) ex-
tends by continuity to F : L
2
(R) L
2
(R
). The inverse Fourier transform F
initially dened on L
1
(R
) L
2
(R
) extends by continuity to F
: L
2
(R
)
L
2
(R). These are unitary operators that preserve L
2
norm and preserve inner
product. Furthermore, F
is the inverse of F.
3.4 Absolute convergence
We have seen that the Fourier transform gives a perfect correspondence between
L
2
(R) and L
2
(R
). For the other spaces the situation is more complex.

The map from a function to its Fourier transform gives a continuous map
from L
1
(R) to part of C
0
(R
). That is, the Fourier transform of an inte-

grable function is continuous and bounded (this is obvious) and approach zero
(Riemann-Lebesgue lemma). Furthermore, this map is one-to-one. That is, the
Fourier transform determines the function.
The inverse Fourier transform gives a continuous map from L
1
(R
) to C
0
(R).
This is also a one-to-one transformation.
One useful fact is that if f is in L
1
(R) and g is in L
2
(R), then the convolution
f g is in L
2
(R). Furthermore,

f g(k) =

f(k) g(k) is the product of a bounded
function with an L
2
(R
) function, and therefore is in L

2
(R
).
However the same pattern of the product of a bounded function with an L
2
function can arise in other ways. For instance, consider the translate f
a
of a
function f in L
2
(R) dened by f
a
(x) = f(xa). Then

f
a
(k) = exp(ika)

f(k).
This is also the product of a bounded function with an L
2
(R
) function.
One can think of this last example as a limiting case of a convolution.
Let
be an approximate function. Then (
)
a
f has Fourier transform
exp(ika)
(k)

f(k). Now let 0. Then (
)
a
f f
a
, while exp(ika)
(k)

f(k)
exp(ika)

f(k).
2
(R) and if f
exists (in the sense that f is an integral

of f) and if f
is also in L
2
(R), then the Fourier transform is in L
1
(R
). As a
consequence f is is C
0
(R).
Proof:

f(k) = (1/
1 +k
2
)
1 +k
2
f(k). Since f is in L
2
(R), it follows
that

f(k) is in L
2
(R). Since f
is in L
2
(R), it follows that k

f(k) is in L
2
(R
).
Hence

1 +k
2
f(k) is in L
2
(R
). Since 1/
1 +k
2
is also in L
2
(R
), it follows
from the Schwarz inequality that

f(k) is in L
1
(R
).
3.5 Fourier transform pairs
There are some famous Fourier transforms.
Fix > 0. Consider the Gaussian
g
(x) =
1
2
2
exp(
x
2
2
2
). (3.16)
Its Fourier transform is
g
(k) = exp(
2
k
2
2
). (3.17)
Proof: Dene the Fourier transform g
(k) by the usual formula. Check that

_
d
dk
+
2
k
_
g
(k) = 0. (3.18)
This proves that
g
(k) = C exp(
2
k
2
2
). (3.19)
Now apply the equality of L
2
norms. This implies that C
2
= 1. By looking at
the case k = 0 it becomes obvious that C = 1.
3.6. PROBLEMS 49
Let > 0. Introduce the Heaviside function H(k) that is 1 for k > 0 and 0
for k < 0. The two basic Fourier transform pairs are
f
(x) =
1
x i
(3.20)
with Fourier transform
(k) = 2iH(k)e
k
. (3.21)
and its complex conjugate
f
(x) =
1
x +i
(3.22)
(k) = 2iH(k)e
k
. (3.23)
These may be checked by computing the inverse Fourier transform. Notice that
f
and its conjugate are not in L

1
(R).
Take 1/ times the imaginary part. This gives the approximate delta func-
tion
(x) =
1
x
2
+
2
. (3.24)
(k) = e
|k|
. (3.25)
Take the real part. This gives the approximate principal value of 1/x function
p
(x) =
x
x
2
+
2
(3.26)
p
(k) = i[H(k)e
k
H(k)e
k
]. (3.27)
3.6 Problems
1. Let f(x) = 1/(2a) for a x a and be zero elsewhere. Find the L
1
(R),
L
2
(R), and L
(R) norms of f, and compare them.

2. Find the Fourier transform of f.
3. Find the L
(R
), L
2
(R
), and L
1
(R
) norms of the Fourier transform, and

compare them.
4. Compare the L
(R
) and L
1
(R) norms for this problem. Compare the
L
(R) and L
1
(R
) norms for this problem.

5. Use the pointwise convergence at x = 0 to evaluate an improper integral.
6. Calculate the convolution of f with itself.
7. Find the Fourier transform of the convolution of f with itself. Verify in
this case that the Fourier transform of the convolution is the product of
the Fourier transforms.
3.7 Poisson summation formula
Theorem: Let f be in L
1
(R) with

f in L
1
(R
) and such that

k
[
f(k)[ < .
Then
2
n
f(2n) =
f(k). (3.28)
Proof: Let
S(t) =
n
f(2n +t). (3.29)
Since S(t) is 2 periodic, we can expand
S(t) =
k
a
k
e
ikt
. (3.30)
It is easy to compute that
a
k
=
1
2
_
2
0
S(t)e
ikt
dt =
1
2
f(k). (3.31)
So the Fourier series of S(t) is absolutely summable. In particular
S(0) =
k
a
k
. (3.32)
3.8 Problems
1. In this problem the Fourier transform is
f(k) =
_

e
ixk
f(x) dx
and the inverse Fourier transform is
f(x) =
_

e
ixk
f(k)
dk
2
.
These provide an isomorphism between the Hilbert spaces L
2
(R, dx) and
L
2
(R,
dk
2
). The norm of f in the rst space is equal to the norm of

f in
the second space. We will be interested in the situation where the Fourier
transform is band-limited, that is, only waves with [k[ a have non-zero
amplitude.
Make the assumption that [k[ > a implies

f(k) = 0. That is, the Fourier
transform of f vanishes outside of the interval [a, a].
Let
g(x) =
sin(ax)
ax
.
3.9. PDE PROBLEMS 51
The problem is to prove that
f(x) =
m=
f(
m
a
)g(x
m
a
).
This says that if you know f at multiples of /a, then you know f at all
points.
Hint: Let g
m
(x) = g(x m/a). The task is to prove that f(x) =
m
c
m
g
m
(x) with c
m
= f(m/a). It helps to use the Fourier transform
of these functions. First prove that the Fourier transform of g(x) is given
by g(k) = /a for [k[ a and g(k) = 0 for [k[ > a. (Actually, it may
be easier to deal with the inverse Fourier transform.) Then prove that
g
m
(k) = exp(imk/a) g(k). Finally, note that the functions g
m
(k) are
orthogonal.
2. In the theory of neural networks one wants to synthesize an arbitrary
function from linear combinations of translates of a xed function. Let f
be a function in L
2
(R). Suppose that the Fourier transform

f(k) ,= 0 for
all k. Dene the translate f
a
by f
a
(x) = f(x a). The task is to show
that the set of all linear combinations of the functions f
a
, where a ranges
over all real numbers, is dense in L
2
(R).
Hint: It is sucient to show that if g is in L
2
(R) with (g, f
a
) = 0 for all
a, then g = 0. (Why is this sucient?) This can be done using Fourier
analysis.
3.9 PDE Problems
1. Consider the initial value problem for the reaction-diusion equation
u
t
=

2
u
x
2
+u u
2
.
We are interested in solutions u with 0 u 1. Find the constant solu-
tions of this equation. Find the linear equations that are the linearizations
about each of these constant solutions.
2. Find the dispersion relations for each of these linear equations. Find the
range of wave numbers (if any) that are responsible for any instability of
these linear equations.
3. Let z = xct and s = t. Find /x and /t in terms of /z and /s.
Write the partial dierential equation in these new variables. A traveling
wave solution is a solution u for which u/s = 0. Write the ordinary
dierential equation for a traveling wave solution (in terms of du/dz).
4. Write this ordinary dierential equation as a rst order system. Take
c > 0. Find the xed points and classify them.
5. Look for a traveling wave solution that goes from 1 at z = to 0 at
z = +. For which values of c are there solutions that remain in the
interval from 0 to 1?
Chapter 4
Complex integration
4.1 Complex number quiz
1. Simplify
1
3+4i
.
2. Simplify [
1
3+4i
[.
3. Find the cube roots of 1.
4. Here are some identities for complex conjugate. Which ones need correc-
tion? z +w = z + w, z w = z w, zw = z w, z/w = z/ w. Make suitable
corrections, perhaps changing an equality to an inequality.
5. Here are some identities for absolute value. Which ones need correction?
[z +w[ = [z[ +[w[, [z w[ = [z[ [w[, [zw[ = [z[[w[, [z/w[ = [z[/[w[. Make
suitable corrections, perhaps changing an equality to an inequality.
6. Dene log(z) so that < log(z) . Discuss the identities e
log(z)
= z
and log(e
w
) = w.
7. Dene z
w
= e
wlog z
. Find i
i
.
8. What is the power series of log(1 +z) about z = 0? What is the radius of
convergence of this power series?
9. What is the power series of cos(z) about z = 0? What is its radius of
convergence?
10. Fix w. How many solutions are there of cos(z) = w with < 'z .
53
54 CHAPTER 4. COMPLEX INTEGRATION
4.2 Complex functions
4.2.1 Closed and exact forms
In the following a region will refer to an open subset of the plane. A dierential
form p dx +q dy is said to be closed in a region R if throughout the region
q
x
=
p
y
. (4.1)
It is said to be exact in a region R if there is a function h dened on the region
with
dh = p dx +q dy. (4.2)
Theorem. An exact form is closed.
The converse is not true. Consider, for instance, the plane minus the origin.
The form (y dx + xdy)/(x
2
+ y
2
) is not exact in this region. It is, however,
exact in the plane minus the negative axis. In this region
y dx +xdy
x
2
+y
2
= d, (4.3)
where /2 < < /2.
Greens theorem. If S is a bounded region with oriented boundary S, then
_
S
p dx +q dy =
_ _
S
(
q
x

p
y
) dxdy. (4.4)
Consider a region R and an oriented curve C in R. Then C 0 (C is
homologous to 0) in R means that there is a bounded region S such that S and
its oriented boundary S are contained in R such that S = C.
Corollary. If p dx +q dy is closed in some region R, and if C 0 in R, then
_
C
p dx +q dy = 0. (4.5)
If C is an oriented curve, then C is the oriented curve with the opposite
orientation. The sum C
1
+ C
2
of two oriented curves is obtained by following
one curve and then the other. The dierence C
1
C
2
is dened by following
one curve and then the other in the reverse direction.
Consider a region R and two oriented curves C
1
and C
2
in R. Then C
1
C
2
(C
1
is homologous to C
2
) in R means that C
1
C
2
0 in R.
Corollary. If p dx + q dy is closed in some region R, and if C
1
C
2
in R,
then
_
C
1
p dx +q dy =
_
C
2
p dx +q dy. (4.6)
4.2. COMPLEX FUNCTIONS 55
4.2.2 Cauchy-Riemann equations
Write z = x +iy. Dene partial dierential operators
z
=

x
+
1
i
y
(4.7)
and
z
=

x

1
i
y
(4.8)
The justication for this denition is the following. Every polynomial in x, y
may be written as a polynomial in z, z, and conversely. Then for each term in
such a polynomial
z
z
m
z
n
= mz
m1
z
n
(4.9)
and
z
z
m
z
n
= z
m
n z
n1
. (4.10)
Let w = u +iv be a function f(z) of z = x +iy. Suppose that this satises
the system of partial dierential equations
w
z
= 0. (4.11)
In this case we say that f(z) is an analytic function of z in this region. Explicitly
(u +iv)
x

(u +iv)
iy
= 0. (4.12)
This gives the Cauchy-Riemann equations
u
x
=
v
y
(4.13)
and
v
x
=
u
y
. (4.14)
4.2.3 The Cauchy integral theorem
Consider an analytic function w = f(z) and the dierential form
wdz = f(z) dz = (u +iv) (dx +idy) = (udx v dy) +i(v dx +udy). (4.15)
According to the Cauchy-Riemann equations, this is a closed form.
Theorem (Cauchy integral theorem) If f(z) is analytic in a region R, and if
C 0 in R, then
_
C
f(z) dz = 0. (4.16)
Example: Consider the dierential form z
m
dz for integer m ,= 1. When
m 0 this is dened in the entire complex plane; when m < 0 it is dened in
the punctured plane (the plane with 0 removed). It is exact, since
z
m
dz =
1
m+ 1
dz
m+1
. (4.17)
On the other hand, the dierential form dz/z is closed but not exact in the
punctured plane.
4.2.4 Polar representation
The exponential function is dened by
exp(z) = e
z
=
n=0
z
n
n!
. (4.18)
It is easy to check that
e
x+iy
= e
x
e
iy
= e
x
(cos(y) +i sin(y)). (4.19)
Sometimes it is useful to represent a complex number in the polar represen-
tation
z = x +iy = r(cos() +i sin()). (4.20)
This can also be written
z = re
i
. (4.21)
From this we derive
dz = dx +i dy = dr e
i
+rie
i
d. (4.22)
This may also be written
dz
z
=
dr
r
+i d. (4.23)
Notice that this does not say that dz/z is exact in the punctured plane. The
reason is that the angle is not dened in this region. However dz/z is exact
in a cut plane, that is, a plane that excludes some line running from the origin
to innity.
Let C(0) be a circle of radius r centered at 0. We conclude that
_
C(0)
f(z) dz =
_
2
0
f(z)z i d. (4.24)
In particular,
_
C(0)
1
z
dz =
_
2
0
i d = 2i. (4.25)
By a change of variable, we conclude that for a circle C(z) of radius r centered
at z we have
_
C(z)
1
z
d = 2i. (4.26)
4.3. COMPLEX INTEGRATION AND RESIDUE CALCULUS 57
4.2.5 Branch cuts
Remember that
dz
z
=
dr
r
+i d (4.27)
is exact in a cut plane. Therefore
dz
z
= d log(z) (4.28)
in a cut plane,
log(z) = log(r) +i (4.29)
Two convenient choices are 0 < < 2 (cut along the positive axis and <
< (cut along the negative axis).
In the same way one can dene such functions as
z = exp(
1
2
log(z)). (4.30)
Again one must make a convention about the cut.
4.3 Complex integration and residue calculus
4.3.1 The Cauchy integral formula
Theorem. (Cauchy integral formula) Let f() be analytic in a region R. Let
C 0 in R, so that C = S, where S is a bounded region contained in R. Let
z be a point in S. Then
f(z) =
1
2i
_
C
f()
z
d. (4.31)
Proof: Let C
(z) be a small circle about z. Let R
be the region R with the

point z removed. Then C C
(z) in R
. It follows that
1
2i
_
C
f()
z
d =
1
2i
_
C
(z)
f()
z
d. (4.32)
It follows that
1
2i
_
C
f()
z
d f(z) =
1
2i
_
C
(z)
f() f(z)
z
d. (4.33)
Consider an arbitrary > 0. The function f() is continuous at = z. Therefore
there is a so small that for on C
(z) the absolute value [f()f(z)[ . Then

the integral on the right hand side has integral with absolute value bounded by
1
2
_
2
0
d = . (4.34)
Therefore the left hand side has absolute value bounded by . Since is arbitrary,
the left hand side is zero.
4.3.2 The residue calculus
Say that f(z) has an isolated singularity at z
0
. Let C
(z
0
) be a circle about z
0
that contains no other singularity. Then the residue of f(z) at z
0
is the integral
res(z
0
) =
1
2i
_
C
(z
0
)
f(z) dz. (4.35)
Theorem. (Residue Theorem) Say that C 0 in R, so that C = S with
the bounded region S contained in R. Suppose that f(z) is analytic in R except
for isolated singularities z
1
, . . . , z
k
in S. Then
_
C
f(z) dz = 2i
k
j=1
res(z
j
). (4.36)
Proof: Let R
be R with the singularities omitted. Consider small circles

C
1
, . . . , C
k
around these singularities such that C C
1
+ +C
k
in R
. Apply
the Cauchy integral theorem to C C
1
. . . C
k
.
If f(z) = g(z)/(z z
0
) with g(z) analytic near z
0
and g(z
0
) ,= 0, then f(z)
is said to have a pole of order 1 at z
0
.
Theorem. If f(z) = g(z)/(z z
0
) has a pole of order 1, then its residue at
that pole is
res(z
0
) = g(z
0
) = lim
zz
0
(z z
0
)f(z). (4.37)
Proof. By the Cauchy integral formula for each suciently small circle C
about z
0
the function g(z) satises
g(z
0
) =
1
2i
_
C
g(z)
z z
0
d =
1
2i
_
C
f(z) dz. (4.38)
This is the residue.
4.3.3 Estimates
Recall that for complex numbers we have [zw[ = [z[[w[ and [z/w[ = [z[/[w[.
Furthermore, we have
[[z[ [w[[ [z w[ [z[ +[w[. (4.39)
When [z[ > [w[ this allows us to estimate
1
[z w[

1
[z[ [w[
. (4.40)
Finally, we have
[e
z
[ = e
z
. (4.41)
4.4. PROBLEMS 59
4.3.4 A residue calculation
Consider the task of computing the integral
_

e
ikx
1
x
2
+ 1
dx (4.42)
where k is real. This is the Fourier transform of a function that is in L
2
and
also in L
1
. The idea is to use the analytic function
f(z) = e
ikz
1
z
2
+ 1
. (4.43)
The rst thing is to analyze the singularities of this function. There are poles
at z = i. Furthermore, there is an essential singularity at z = .
First look at the case when k 0. The essential singularity of the exponen-
tial function has the remarkable feature that for z 0 the absolute value of
e
ikz
is bounded by one. This suggests looking at a closed oriented curve C
a
in
the upper half plane. Take C
a
to run along the x axis from a to a and then
along the semicircle z = ae
i
from = 0 to = . If a > 1 there is a singularity
at z = i inside the curve. So the residue is the value of g(z) = e
ikz
/(z +i) at
z = i, that is, g(i) = e
k
/(2i). By the residue theorem
_
C
a
e
ikz
1
z
2
+ 1
dz = e
k
(4.44)
for each a > 1.
Now let a . The contribution from the semicircle is bounded by
_

0
1
a
2
1
a d =
a
a
2
1
. (4.45)
We conclude that for k 0
_

e
ikx
1
x
2
+ 1
dx = e
k
. (4.46)
Next look at the case when k 0. In this case we could look at an oriented
curve in the lower half plane. The integral runs from a to a and then around
a semicircle in the lower half plane. The residue at z = i is e
k
/(2i). By
the residue theorem, the integral is e
k
. We conclude that for all real k
_

e
ikx
1
x
2
+ 1
dx = e
|k|
. (4.47)
4.4 Problems
1. Evaluate
_

0
1
x(4 +x
2
)
dx
by contour integration. Show all steps, including estimation of integrals
that vanish in the limit of large contours.
2. In the following problems f(z) is analytic in some region. We say that
f(z) has a root of multiplicity m at z
0
if f(z) = (z z
0
)
m
h(z), where h(z)
is analytic with h(z
0
) ,= 0. Find the residue of f
(z)/f(z) at such a z
0
.
3. Say that f(z) has several roots inside the contour C. Evaluate
1
2i
_
C
f
(z)
f(z)
dz.
4. Say that
f(z) = a
0
+a
1
z +a
2
z
2
+ +a
n
z
n
is a polynomial. Furthermore, suppose that C is a contour surrounding
the origin on which
[a
k
z
k
[ > [f(z) a
k
z
k
[.
Show that on this contour
f(z) = a
k
z
k
g(z)
where
[g(z) 1[ < 1
on the contour. Use the result of the previous problem to show that the
number of roots (counting multiplicity) inside C is k.
5. Find the number of roots (counting multiplicity) of z
6
+3z
5
+1 inside the
unit circle.
4.5 More residue calculus
4.5.1 Jordans lemma
Jordans lemma says that for b > 0 we have
1
_

0
e
b sin()
d
1
b
. (4.48)
To prove it, it is sucient to estimate twice the integral over the interval
from 0 to /2. On this interval use the inequality (2/) sin(). This gives
2
_
2
0
e
2b/
d =
1
b
(1 e
b
)
1
b
. (4.49)
4.5. MORE RESIDUE CALCULUS 61
4.5.2 A more delicate residue calculation
Consider the task of computing the integral
_

e
ikx
1
x i
dx (4.50)
2
but not
in L
1
f(z) = e
ikz
1
z i
. (4.51)
The rst thing is to analyze the singularities of this function. There is a pole
First look at the case when k < 0. Take C
a
to run along the x axis from a
to a and then along the semicircle z = ae
i
in the upper half plane from = 0
to = . If a > 1 there is a singularity at z = i inside the curve. So the residue
is the value of g(z) = e
ikz
at z = i, that is, g(i) = e
k
. By the residue theorem
_
C
a
e
ikz
1
z i
dz = 2ie
k
(4.52)
for each a > 1.
Now let a . The contribution from the semicircle is bounded using
Jordans lemma:
_

0
e
ka sin()
1
a 1
a d
1
ka
a
a 1
(4.53)
We conclude that for k < 0
_

e
ikx
1
x i
dx = 2ie
k
. (4.54)
Next look at the case when k < 0. In this case we could look at an oriented
a semicircle in the lower half plane. The residue is zero. We conclude using
Jordans lemma that for k < 0
_

e
ikx
1
x i
dx = 0. (4.55)
4.5.3 Cauchy formula for derivatives
Theorem. (Cauchy formula for derivatives) Let f() be analytic in a region
R including a point z. Let C be an oriented curve in R such that for each
suciently small circle C(z) about z, C C(z) in R. Then the mth derivative
satises
1
m!
f
(m)
(z) =
1
2i
_
C
f()
( z)
m+1
d. (4.56)
Proof: Dierentiate the Cauchy integral formula with respect to z a total of
m times.
4.5.4 Poles of higher order
If f(z) = g(z)/(z z
0
)
m
with g(z) analytic near z
0
and g(z
0
) ,= 0 and m 1,
then f(z) is said to have a pole of order m at z
0
.
Theorem. If f(z) = g(z)/(z z
0
)
m
has a pole of order m, then its residue
at that pole is
res(z
0
) =
1
(m1)!
g
(m1)
(z
0
). (4.57)
Proof. By the Cauchy formula for derivatives for each suciently small circle
C about z the m1th derivative satises
1
(m1)!
g
(m1)
(z
0
) =
1
2i
_
C
g(z)
(z z
0
)
m
dz =
1
2i
_
C
f(z) dz. (4.58)
The expression given in the theorem is evaluated in practice by using the
fact that g(z) = (z z
0
)
m
f(z) for z near z
0
, performing the dierentiation, and
then setting z = z
0
. This is routine, but it can be tedious.
4.5.5 A residue calculation with a double pole
Conder the task of computing the integral
_

e
ikx
1
(x
2
+ 1)
2
dx (4.59)
2
and
also in L
1
f(z) = e
ikz
1
(z
2
+ 1)
2
. (4.60)
The rst thing is to analyze the singularities of this function. There are poles
First look at the case when k 0. Consider a closed oriented curve C
a
in the
upper half plane. Take C
a
to run along the x axis from a to a and then along
the semicircle z = ae
i
from = 0 to = . If a > 1 there is a singularity at
z = i inside the curve. The pole there is of order 2. So the residue is calculated
by letting g(z) = e
ikz
/(z +i)
2
, taking the derivative, and evaluating at z = i,
that is, g
(i) = (1 k)e
k
/(4i). By the residue theorem
_
C
a
e
ikz
1
z
2
+ 1
dz =
1
2
(1 k)e
k
(4.61)
for each a > 1.
Now let a . The contribution from the semicircle vanishes in that limit.
We conclude that for k 0
_

e
ikx
1
(x
2
+ 1)
2
dx =
1
2
(1 k)e
k
. (4.62)
4.6. THE TAYLOR EXPANSION 63
Next look at the case when k 0. In this case we could look at an oriented
a semicircle in the lower half plane. The residue at z = i is (1 +k)e
k
/(4i).
By the residue theorem, the integral is (1 + k)e
k
/2. We conclude that for
all real k
_

e
ikx
1
(x
2
+ 1)
2
dx =
1
2
(1 +[k[)e
|k|
(4.63)
4.6 The Taylor expansion
4.6.1 Radius of convergence
Theorem. Let f(z) be analytic in a region R including a point z
0
. Let C(z
0
) be
a circle centered at z
0
such that C(z
0
) and its interior are in R. Then for z in
the interior of C(z
0
)
f(z) =
m=0
1
m!
f
(m)
(z
0
)(z z
0
)
m
. (4.64)
Proof. For each xed the function of z given by
1
z
=
1
z
0
+z
0
z
=
1
( z
0
)
1
1
zz
0
z
0
=
1
( z
0
)
m=0
_
z z
0
z
0
_
m
.
(4.65)
has a geometric series expansion. Multiply by (1/2i)f() d and integrate
around C(z
0
). On the left hand side apply the Cauchy integral formula to get
f(z).
In each term in the expansion on the right hand side apply the Cauchy
formula for derivatives in the form
1
2i
_
C(z
0
)
f()
( z
0
)
m+1
d =
1
m!
f
(m)
(z
0
). (4.66)
This theorem is remarkable because it shows that the condition of analyticity
implies that the Taylor series always converges. Furthermore, take the radius
of the circle C(z
0
) as large as possible. The only constraint is that there must
be a function that is analytic inside the circle and that extends f(z). Thus one
must avoid the singularity of this extension of f(z) that is closest to z
0
. This
explains why the radius of convergence of the series is the distance from z
0
to
this nearest singularity.
The reason that we talk of an analytic extension is that articialities in the
denition of the function, such as branch cuts, should not matter. On the other
hand, singularities such as poles, essential singularities, and branch points are
intrinsic. The radius of convergence is the distance to the nearest such intrinsic
singularity.
If one knows an analytic function near some point, then one knows it all
the way out to the radius of convergence of the Taylor series about that point.
But then for each point in this larger region there is another Taylor series. The
function is then dened out to the radius of convergence associated with that
point. The process may be carried out over a larger and larger region, until
blocked by some instrinsic singularity. It is known as analytic continuation.
If the analytic continuation process begins at some point and winds around a
branch point, then it may lead to a new denition of the analytic function at the
original point. This appears to lead to the necessity of introducing an artical
branch cut in the denition of the function. However this may be avoided by
introducing the concept of Riemann surface.
4.6.2 Riemann surfaces
When an analytic function has a branch cut, it is an indicator of the fact that
the function should not be thought of not as a function on a region of the
complex plane, but instead as a function on a Riemann surface. A Riemann
surface corresponds to several copies of a region in the complex plane. These
copies are called sheets. Where one makes the transition from one sheet to
another depends on the choice of branch cut. But the Riemann surface itself is
independent of the notion of sheet.
As an example, take the function w =
z. The most natural denition of this

function is as a function on the curve w
2
= z. This is a kind of two-dimensional
parabola in a four dimensional space of two complex variables. The value of the
square root function on the point with coordinates z, w on the parabola w
2
= z
is w. Notice that if z is given, then there are usually two corresponding values
of w. Thus if we want to think of

z as a function of a complex variable, then
it is ambiguous. But if we think of it as a function on the Riemann surface,
then it is perfectly well dened. The Riemann surface has two sheets. If we
wish, we may think of sheet I as the complex plane cut along the negative axis.
Sheet II is another copy of the complex plane cut along the negative axis. As
z crosses the cut, it changes from one sheet to the other. The value of w also
varies continuously, and it keeps track of what sheet the z is on.
As another example, take the function w =
_
z(z 1). The most natural
denition of this function is as a function on the curve w
2
= z(z 1). This
is a kind of two-dimensional circle in a four dimensional space of two complex
variables. The value of the function on the point with coordinates z, w on the
circle w
2
= z(z 1) is w. Notice that if z is given, then there are usually two
corresponding values of w. Thus if we want to think of
_
z(z 1) as a function
of a complex variable, then it is ambiguous. But if we think of it as a function
on the Riemann surface, then it is perfectly well dened. The Riemann surface
has two sheets. If we wish, we may think of sheet I as the complex plane cut
between 0 and 1. Sheet II is another copy of the complex plane, also cut between
0 and 1. As z crosses the cut, it changes from one sheet to the other. The value
of w also varies continuously, and it keeps track of what sheet the z is on.
As a nal example, take the function w = log z. The most natural denition
of this function is as a function on the curve exp(w) = z. This is a two-
dimensional curve in a four dimensional space of two complex variables. The
4.6. THE TAYLOR EXPANSION 65
value of the logarithm function on the point with coordinates z, w on the curve
exp(w) = z is w. Notice that if z is given, then there are innitely many
corresponding values of w. Thus if we want to think of log z as a function of a
complex variable, then it is ambiguous. But if we think of it as a function on
the Riemann surface, then it is perfectly well dened. The Riemann surface has
innitely many sheets. If we wish, we may think of each sheet as the complex
plane cut along the negative axis. As z crosses the cut, it changes from one
sheet to the other. The value of w also varies continuously, and it keeps track
of what sheet the z is on. The innitely many sheets form a kind of spiral that
winds around the origin innitely many times.
Chapter 5
Distributions
5.1 Properties of distributions
Consider the space C
c
(R) of complex test functions. These are complex func-
tions dened on the real line, innitely dierentiable and with compact support.
A distribution is a linear functional from this space to the complex numbers. It
must satisfy a certain continuity condition, but we shall ignore that. The value
of the distribution F on the test function is written F, ).
If f is a locally integrable function, then we may dene a distribution
F, ) =
_

f(x)(x) dx. (5.1)

Thus many functions dene distributions. This is why distributions are also
called generalized functions.
A sequence of distributions F
n
is said to converge to a distribution F if for
each test function the numbers F
n
, ) converge to the number F, ).
Example. The distribution
a
is dened by
a
, ) = (a). (5.2)
This is not given by a locally integrable function. However a distribution may
be written as a limit of functions. For instance, let > 0 and consider the
function
(x a) =
1
(x a)
2
+
2
. (5.3)
The limit of the distributions dened by these locally integrable functions as
0 is
a
. For this reason the distribution is often written in the incorrect but
suggestive notation
a
, ) =
_

(x a)(x) dx. (5.4)

Operations on distributions are dened by looking at the example of a dis-
tribution dened by a function and applying the integration by parts formula
67
68 CHAPTER 5. DISTRIBUTIONS
in that case. Thus, for instance, the derivative is dened by
F
, ) = F,
). (5.5)
Example: Consider the locally integrable Heaviside function given by H(x) =
1 for x > 0, H(x) = 0 for x < 0. Then H
= . Here is given by
, ) = (0). (5.6)
Example: Consider the locally integrable log function f(x) = log [x[. Then
f
(x) = PV 1/x. Here the principal value PV 1/x is given by

PV
1
x
, ) = lim
0
_

x
x
2
+
2
(x) dx.. (5.7)
This can be seen by writing log [x[ = lim
log
x
2
+
2
.
A distribution can be approximated by more than one sequence of functions.
For instance, for each a > 0 let log
a
([x[) = log([x[) for [x[ a and log
a
([x[) =
log(a) for [x[ < a. Then log
a
([x[) also approaches log([x[) as a 0. So its
derivative is another approximating sequence for PV 1/x. This says that
PV
1
x
, ) = lim
a0
_
|x|>a
1
x
(x) dx. (5.8)
We can use this sequence to compute the derivative of PV 1/x. We get
_

_
d
dx
PV
1
x
_
(x) dx = lim
a0
_
|x|>a
1
x
(x) dx = lim
a0
_
_
|x|>a
1
x
2
(x) dx +
(a) +(a)
a
_
.
(5.9)
But [(a) +(a) 2(0)]/a converges to zero, so this is
_

_
d
dx
PV
1
x
_
(x) dx = lim
a0
_
|x|>a
1
x
2
[(x) (0)] dx. (5.10)
Another operator is multiplication of a distribution by a smooth function g
in C
. This is dened in the obvious way by

g F, ) = F, g). (5.11)
Distributions are not functions. They may not have values at points, and
in general nonlinear operations are not dened. For instance, the square of a
distribution is not always a well dened distribution.
Also, some algebraic operations involving distributions are quite tricky. Con-
sider, for instance, the associative law. Apply this to the three distributions
(x), x, and PV 1/x. Clearly the product (x) x = 0. On the other hand, the
product x PV 1/x = 1 is one. So if the associate law were to hold, we would
get
0 = 0 PV
1
x
= ((x) x) PV
1
x
= (x) (x PV
1
x
) = (x) 1 = (x). (5.12)
5.2. MAPPING DISTRIBUTIONS 69
5.2 Mapping distributions
A test function is naturally viewed as a covariant object, so the distribution
F is contravariant. A proper function is a function such that the inverse image
of each compact set is compact. It is natural to dene the forward push of a
distribution F by a smooth proper function g by g[F], ) = F, g). Example:
If F is the distribution (x3) and if u = g(x) = x
2
4, then the forward push
is (u 5). This is because
_
(u 5)(u) du =
_
(x 3)(x
2
4) dx.
On the other hand, it is actually more common to think of a distribution
as being a covariant object, since a distribution is supposed to be a generalized
function. The backward pull of the distribution by a smooth function g is
dened in at least some circumstances by
F g, ) = F, g[]). (5.13)
Here
g[](u) =
g(x)=u
(x)
[g
(x)[
. (5.14)
Example. Let u = g(x) = x
2
4, with a > 0. Then the backward pull of
(u) under g is (x
2
4) = (1/4)((x 2) +(x + 2)). This is because
g[](u) =
(
u
2
+ 4) +(
u
2
+ 4)
2
u
2
+ 4
. (5.15)
So if F = , then
F g =
1
4
(
2
+
2
) (5.16)
Example: The backward pull is not always dened. To consider a dis-
tribution as a covariant object is a somewhat awkward act in general. Let
u = h(x) = x
2
. The backward pull of (u) by h is (x
2
), which is not dened.
Example: The general formula for the pull back of the delta function is
(g(x)) =
g(a)=0
1
[g
(a)[
(x a). (5.17)
The most important distributions are , PV 1/x, 1/(xi0), and 1/(x+i0).
These are the limits of the functions
(x), x/(x
2
+
2
), 1/(x i), 1/(x +i) as
0. The relations between these functions are given by
(x) =
1
x
2
+
2
=
1
2i
_
1
x i

1
x +i
_
. (5.18)
and
x
x
2
+
2
=
1
2
_
1
x i
+
1
x +i
_
. (5.19)
5.3 Radon measures
A Radon measure is a positive distribution. That is, it is a linear functional
on C
c
(R) such that for each test function the condition 0 implies
that the value , ) 0. Every Radon measure extends uniquely by continuity
to a linear function on C
c
(R), the space of continuous functions with compact
support. Each positive locally integrable function h denes a Radon measure by
integrating (x) times h(x) dx. Also, the point mass distributions
a
are Radon
measures.
It is common to write the value of a Radon measure in the form
, ) =
_
(x) d(x). (5.20)
What is remarkable is that the theory of Lebesgue integration works for Radon
measures. That is, given a real function f 0 that is only required to be Borel
measurable, there is a natural denition of the integral such that
0
_
f(x) d(x) +. (5.21)
Furthermore, if f is a complex Borel measurable function such that [f[ has nite
integral, then the integral of f is dened and satises.
[
_
f(x) d(x)[
_
[f(x)[ d(x) < +. (5.22)
5.4 Approximate delta functions
It might seem that one could replace the notion of distribution by the notion of
a sequence of approximating functions. This is true in some sense, but the fact
is that many dierent sequences may approximate the same distribution. Here
is a result of that nature.
Theorem. Let
1
(u) 0 be a positive function with integral 1. For each > 0
dene
(x) = (x/)/. Then the functions
converge to the distribution as

tends to zero.
The convergence takes place in the sense of distributions (smooth test func-
tions with compact support) or even in the sense of Radon measures (continuous
test functions with compact support). Notice that there are no continuity or
symmetry assumptions on the initial function.
The proof of this theorem is simple. Each
has integral one. Consider a

bounded continuous function . Then
_

(x)(x) dx =
_

1
(u)(u) du. (5.23)
The dominated convergence theorem shows that this approaches the integral
_

1
(u)(0) du = (0). (5.24)
5.5. PROBLEMS 71
Here is an even stronger result.
Theorem. For each > 0 let
0 be a positive function with integral 1.

Suppose that for each a > 0 that
_
|x|>a
(x) dx 0 (5.25)
as 0. Then the functions
converge to the distribution as tends to

zero.
Here is a proof of this more general theorem. Let H
(a) =
_
a
0

(x) dx. Then

for each a < 0 we have H
(a) 0, and for each a > 0 we have 1 H
(a) 0.
In other words, for each a ,= 0 we have H
(a) H(a) as 0. Since the

functions H
are uniformly bounded, it follows from the dominated convergence

theorem that H
H in the sense of distributions. It follows by dierentiation

that
in the sense of distributions.

5.5 Problems
If F and G are distributions, and if at least one of them has compact support,
then their convolution F G is dened by
F G, ) = F
x
G
y
, (x +y)).
This product is commutative. It is also associative if at least two of the three
factors have compact support.
1. If F and G are given by locally integrable functions f and g, and at
least one has compact support, then F G is given by a locally integrable
function
(f g)(z) =
_

f(x)g(z x) dx =
_

f(z y)g(y) dy.

2. If G is given by a test function g, then F g is given by a smooth function
(F g)(z) = F
x
, g(z x)).
3. Calculate the convolution 1
.
4. Calculate the convolution
H, where H is the Heaviside function.

5. Calculate the convolution (1
) H and also calculate the convolution

1(
H). What does this say about the associative law for convolution?
6. Let L be a constant coecient linear dierential operator. Let u be a
distribution that is a fundamental solution, that is, let Lu = . Let G be
a distribution with compact support. Show that the convolution F = uG
satises the equation LF = G. Hint: Write LF, ) = F, L
), where L
is adjoint to L.
7. Take L = d
2
/dx
2
. Is there a fundamental solution that has support in
a bounded interval? Is there a fundamental solution that has support in
a semi-innite interval?
5.6 Tempered distributions
Let d/dx be the operator of dierentiation, and let x be the operator of mul-
tiplication by the coordinate x. The space o of rapidly decreasing smooth test
functions consists of the functions in L
2
such that every nite product of the
operators d/dx and x applied to is also in L
2
. The advantage of this denition
is that the space o is clearly invariant under Fourier transformation.
A tempered distribution is a linear functional on o satisfying the appro-
priate continuity property. Every tempered distribution restricts to dene a
distribution. So tempered distributions are more special.
The advantage of tempered distributions is that one can dene Fourier trans-
forms

F of tempered distributions F. The denition is
F, ) = F,

). (5.26)
Here

is the Fourier transform of .
Here are some Fourier transforms for functions. The rst two are easy. They
are
_

e
ikx
1
x i
dx = 2ie
k
H(k) (5.27)
and
_

e
ikx
1
x +i
dx = 2ie
k
H(k). (5.28)
Subtract these and divide by 2i. This gives
_

e
ikx
(x) dx = e
|k|
. (5.29)
Instead, add these and divide by 2. This gives
_

e
ikx
x
x
2
+
2
dx = ie
|k|
sign(k). (5.30)
The corresponding Fourier transforms for distributions are
F[1/(x i0)] = 2iH(k) (5.31)
and
F[1/(x +i0)] = 2iH(k). (5.32)
Also,
F[(x)] = 1 (5.33)
5.6. TEMPERED DISTRIBUTIONS 73
and
F[PV
1
x
] = isign(k). (5.34)
Example: Here is a more complicated calculation. The derivative of PV 1/x
is the distribution
d
dx
PV
1
x
=
1
x
2
+c(x), (5.35)
where c is the innite constant
c =
_

1
x
2
dx. (5.36)
This makes rigorous sense if interprets it as
_

_
d
dx
PV
1
x
_
(x) dx = lim
a0
_
|x|>a
1
x
2
[(x) (0)] dx. (5.37)
One can get an intuitive picture of this result by graphing the approximating
functions. The key formula is
d
dx
x
x
2
+
2
=
x
2
(x
2
+
2
)
2
+c
2
3
(x
2
+
2
)
2
, (5.38)
where c
= /(2). This is an approximation to 1/x

2
plus a big constant times
an approximation to the delta function.
The Fourier transform of the derivative is obtained by multiplying the Fourier
transform of PV 1/x by ik. Thus the Fourier transform of 1/x
2
+ c(x) is ik
times isign(k) which is [k[.
This example is interesting, because it looks at rst glance that the derivative
of PV 1/x should be 1/x
2
, which is negative denite. But the correct answer
for the derivative is 1/x
2
+ c(x), which is actually positive denite. And in
fact its Fourier transform is positive.
For each of these formula there is a corresponding inverse Fourier transform.
For, instance, the inverse Fourier transform of 1 is
(x) =
_

e
ikx
dk
2
=
_

0
cos(kx)
dk
2
. (5.39)
Of course such an equation is interpreted by integrating both sides with a test
function.
Another formula of the same type is gotten by taking the inverse Fourier
transform of isign(k). This is
PV
1
x
= i
_

e
ikx
sign(k)
dk
2
=
_

0
sin(kx) dk. (5.40)
5.7 Poisson equation
We begin the study of fundamental solutions of partial dierential equations.
These are solutions of the equation Lu = , where L is the dierential operator,
and is a point source.
Let us start with the equation in one space dimension:
_
d
2
dx
2
+m
2
_
u = (x). (5.41)
This is an equilibrium equation that balances a source with losses due to diu-
sion and to dissipation (when m > 0). Fourier transform. This gives
_
k
2
+m
2
_
u(k) = 1. (5.42)
The solution is
u(k) =
1
k
2
+m
2
. (5.43)
There is no problem of division by zero. The inverse Fourier transform is
u(x) =
1
2m
e
m|x|
. (5.44)
This is the only solution that is a tempered distribution. (The solutions of the
homogeneous equation all grow exponentially.)
What happens when m = 0? This is more subtle. The equation is
d
2
dx
2
u = (x). (5.45)
Fourier transform. This gives
k
2
u(k) = 1. (5.46)
Now there is a real question about division by zero. Furthermore, the homo-
geneous equation has solutions that are tempered distributions, namely linear
combinations of (k) and
(k). The nal result is that the inhomogeneous equa-

tion does have a tempered distribution solution, but it needs careful denition.
The solution is
u(k) =
1
k
2
+(k). (5.47)
This may be thought of as the derivative of PV 1/k. The inverse Fourier
transform of PV 1/k is (1/2)isign(x). So the inverse Fourier transform of
d/dkPV 1/k is (ix)(1/2)isign(x) = (1/2)[x[. Thus
u(x) =
1
2
[x[ (5.48)
is a solution of the inhomogeneous equation. The solutions of the homogeneous
equation are linear combinations of 1 and x. None of these solutions are a
5.8. DIFFUSION EQUATION 75
good description of diusive equilibrium. In fact, in one dimension there is no
diusive equilibrium.
The next case that is simple to compute and of practical importance is the
equation in dimension 3. This is
_
2
+m
2
_
u = (x). (5.49)
This is an equilibrium equation that balances a source with losses due to diu-
sion and to dissipation (when m > 0). Fourier transform. This gives
_
k
2
+m
2
_
u(k) = 1. (5.50)
The solution is
u(k) =
1
k
2
+m
2
. (5.51)
The inverse Fourier transform in the three dimension case may be computed by
going to polar coordinates. It is
u(x) =
1
4[x[
e
m|x|
. (5.52)
What happens when m = 0? The situation is very dierent in three dimen-
sions. The equation is
2
u = (x). (5.53)
Fourier transform. This gives
k
2
u(k) = 1. (5.54)
The inhomogeneous equation has a solution
u(k) =
1
k
2
. (5.55)
But now this is a locally integrable function. It denes a tempered distribution
without any regularization. Thus
u(x) =
1
4[x[
(5.56)
is a solution of the inhomogeneous equation. In three dimensions there is dif-
fusive equilibrium. There is so much room that the eect of the source can be
completely compensated by diusion alone.
5.8 Diusion equation
The diusion equation or heat equation is
_

t

1
2
2

2
x
2
_
u = (x)(t). (5.57)
It says that the time rate of change is entirely due to diusion. Fourier trans-
form. We get
(i +
1
2
2
k
2
) u(k, ) = 1. (5.58)
This has solution
u(k, ) =
1
i +
1
2
2
k
2
=
1
i
1
i
1
2
2
k
2
. (5.59)
Here the only division by zero is when both and k are zero. But this is not
so serious, because it is clear how to regularize. We can use the fact that the
inverse Fourier transform of 1/( i) with > 0 is iH(t)e
t
. So we have
u(k, t) = H(t)e
2
tk
2
2
. (5.60)
This is a Gaussian, so the fundamental solution is
u(x, t) = H(t)
1
2
2
t
e
x
2
2
2
t
. (5.61)
5.9 Wave equation
We will look for the solution of the wave equation with a point source at time
zero that lies in the forward light cone. The wave equation in 1 + 1 dimensions
is
_

2
t
2
c
2

2
x
2
_
u = (x)(t). (5.62)
Fourier transform. We get
(
2
+c
2
k
2
) u(k, ) = 1. (5.63)
This has solution
u(k, ) =
1
( i0)
2
c
2
k
2
. (5.64)
The division by zero is serious, but it is possible to regularize. The choice
of regularization is not the only possible one, but we shall see that it is the
convention that gives propagation into the future. We can write this also as
u(k, ) =
1
2c[k[
_
1
i0 c[k[

1
i0 +c[k[
_
. (5.65)
We can use the fact that the inverse Fourier transform of 1/( i0) is iH(t).
So we have
u(k, t) =
1
2c[k[
iH(t)
_
e
ic|k|t
e
ic|k|t
_
=
1
c[k[
sin(c[k[t)H(t). (5.66)
5.10. HOMOGENEOUS SOLUTIONS OF THE WAVE EQUATION 77
It is easy to check that this is the Fourier transform of
u(x, t) =
1
2c
[H(x +ct) H(x ct)] H(t). (5.67)
The wave equation in 3 + 1 dimensions is
_

2
t
2
c
2
2
_
u = (x)(t). (5.68)
Fourier transform. We get
(
2
+c
2
k
2
) u(k, ) = 1. (5.69)
This has solution
u(k, ) =
1
( i0)
2
c
2
k
2
. (5.70)
Again we can write this as
u(k, ) =
1
2c[k[
_
1
i0 c[k[

1
i0 +c[k[
_
. (5.71)
Thus we have again
u(k, t) =
1
2c[k[
iH(t)
_
e
ic|k|t
e
ic|k|t
_
=
1
c[k[
sin(c[k[t)H(t). (5.72)
However the dierence is that the k variable is three dimensional. It is easy to
check that this is the Fourier transform of
u(x, t) =
1
4c[x[
([x[ ct)H(t). (5.73)
This is a beautiful formula. It represents an expanding sphere of inuence, going
into the future. Inside the sphere it is dark. The solution has an even more
beautiful expression that exhibits the symmetry:
u(x, t) =
1
2c
(x
2
c
2
t
2
)H(t). (5.74)
5.10 Homogeneous solutions of the wave equa-
tion
The polynomial
2
k
2
vanishes on an entire cone, so it is not surprising that
the wave equation has a number of interesting homogeneous solutions. The
most important ones are
u(k, t) =
1
c[k[
sin(c[k[t) (5.75)
and its derivative
v(k, t) = cos(c[k[t). (5.76)
These are the solutions that are used in constructing solutions of the initial
value problem.
It is interesting to see what these solutions look like in the frequency repre-
sentation. The result is
u(k, ) = i
1
c[k[
[( c[k[) ( +c[k[)] = 2i(
2
c
2
k
2
)sign() (5.77)
and
v(k, ) = i u(k, ) = [( c[k[) +( +c[k[)] = 2[[(
2
c
2
k
2
). (5.78)
5.11 Problems
1. Show that
1
x
1
3
is a locally integrable function and thus denes a distribu-
tion. Show that its distribution derivative is
d
dx
1
x
1
3
=
1
3
1
x
4
3
+c(x), (5.79)
where
c =
1
3
_

1
x
4
3
dx (5.80)
Hint: To make this rigorous, consider x/(x
2
+
2
)
2
3
.
2. Show that
1
|x|
1
3
is a locally integrable function and thus denes a distrib-
ution. Show that its distribution derivative is
d
dx
1
x
1
3
=
1
3
1
x
4
3
sign(x). (5.81)
The right hand side is not locally integrable. Explain the denition of the
right hand side as a distribution. Hint: To make this rigorous, consider
1/(x
2
+
2
)
1
6
.
3. Discuss the contrast between the results in the last two problems. It
may help to draw some graphs of the functions that approximate these
distributions.
4. Let m > 0. Use Fourier transforms to nd a tempered distribution u that
is the fundamental solution of the equation
d
2
dx
2
u +m
2
u = (x). (5.82)
Is this a unique tempered distribution solution? Explain.
5.12. ANSWERS TO FIRST TWO PROBLEMS 79
5. Let m > 0. Consider Euclidian space of dimension n = 3. Use Fourier
transforms to nd a tempered distribution u that is the fundamental so-
lution of the equation
2
u +m
2
u = (x). (5.83)
5.12 Answers to rst two problems
1. The function
1
x
1
3
is locally integrable, while its pointwise derivative
1
3
1
x
4
3
is not. But we do have
d
dx
x
(x
2
+
2
)
2
3
=
1
3
x
2
(x
2
+
2
)
5
3
+

2
(x
2
+
2
)
5
3
. (5.84)
Let
c
=
_

1
3
x
2
(x
2
+
2
)
5
3
dx (5.85)
which is easily seen to be proportional to 1/
1
3
. From the fundamental
theory of calculus
_

2
(x
2
+
2
)
5
3
dx = c
. (5.86)
Write
(x) =
1
c
2
(x
2
+
2
)
5
3
. (5.87)
Then
(x) (x) in the sense of distributions. Furthermore,

_

d
dx
1
x
1
3
(x) dx = lim
0
_

[
1
3
x
2
(x
2
+
2
)
5
3
+c
(x)](x) dx. (5.88)

Since for smooth enough
_
c
(x)[(x) (0)] dx 0, (5.89)

this may be written in the simple form
_

d
dx
1
x
1
3
(x) dx =
_

1
3
1
x
4
3
[(x) (0)] dx. (5.90)
Notice that the integrand is integrable for each test function .
2. The function
1
|x|
1
3
is locally integrable, while its pointwise derivative
1
3
1
x
4
3
sign(x)
is not. We have
d
dx
1
(x
2
+
2
)
1
6
=
1
3
x
(x
2
+
2
)
7
6
. (5.91)
Since this derivative is an odd function, we have
_

d
dx
1
[x[
1
3
(x) dx = lim
0
_

1
3
x
(x
2
+
2
)
7
6
[(x) (0)] dx. (5.92)
We can write this as
_

d
dx
1
[x[
1
3
(x) dx =
_

1
3
1
x
4
3
sign(x)[(x) (0)] dx. (5.93)
Again the integrand is integrable.
Chapter 6
Bounded Operators
6.1 Introduction
This chapter deals with bounded linear operators. These are bounded linear
transformations of a Hilbert space into itself. In fact, the chapter treats four
classes of operators: nite rank, Hilbert-Schmidt, compact, and bounded. Every
nite rank operator is Hilbert-Schmidt. Every Hilbert-Schmidt operator is com-
pact. Every compact operator is bounded.
We shall see in the next chapter that it is also valuable to look at an even
broader class of operators, those that are closed and densely dened. Every
bounded operator (everywhere dened) is closed and densely dened. However
the present chapter treats only bounded operators.
6.2 Bounded linear operators
Let H be a Hilbert space (a vector space with an inner product that is a complete
metric space). A linear transformation K : H H is said to be bounded if
it maps bounded sets into bounded sets. This is equivalent to there being a
constant M with
|Ku| M|u|. (6.1)
Let M
2
be the least such M. This is called uniform norm of K and is written
|K|
or simply |K| when the context makes this clear. [The subscript in M
2
is supposed to remind us that we are dealing with Hilbert spaces like L
2
. The
subscript in |K|
, on the other hand, tells that we are looking at a least upper

bound.]
If K and L are bounded operators, their sum K + L and product KL are
bounded. Furthermore, we have |K + L|
|K|
+ |L|
and |KL|

|K|
|L|
.
A bounded operator K always has an adjoint K
that is a bounded operator.

81
82 CHAPTER 6. BOUNDED OPERATORS
It is the unique operator with the property that
(u, K
v) = (Ku, v) (6.2)
for all u, v in H. Furthermore K
= K, and |K
= |K|
. For products
we have (KL)
= L
, in the opposite order.

It is also possible to dene the adjoint of an operator from one Hilbert space
to another. Here is a special case. Let g be a vector in the Hilbert space H.
Then g denes a linear transformation from C to H by sending z to the vector
zg. The adjoint is a linear transformation from H to C, denoted by g
. It is the
transformation that sends v to (g, v), so g
v = (g, v). The adjointness relation

is z g
v = (zg, v) which is just z(g, v) = (zg, v).

Let f be another vector. Dene the operator K from H to H by Ku =
f(g, u), that is, K = fg
. Then the adjoint of K is K
= gf
.
Example: Hilbert-Schmidt integral operators. Let
(Kf)(x) =
_
k(x, y)f(y) dy. (6.3)
with k in L
2
, that is,
|k|
2
2
=
_ _
[k(x, y)[
2
dxdy < . (6.4)
Then K is bounded with norm |K|
|k|
2
.
Proof: Fix x. Apply the Schwarz inequality to the integral over y in the def-
inition of the operator. This gives that [(Kf)(x)[
2
_
[k(x, y)[
2
dy
_
[f(y)[
2
dy.
Now integrate over x.
Example: Interpolation integral operators. Let K be an integral operator
such that
M
1
= sup
y
_
[k(x, y)[ dx < (6.5)
and
M
= sup
x
_
[k(x, y)[ dy < . (6.6)
(Here we choose to think of k(x, y) as a function of x and y, not as a Schwartz
distribution like (x y).) Let 1/p +1/q = 1. Then for each p with 1 p
the norm M
p
of K as an operator on L
p
is bounded by M
p
M
1
p
1
M
1
q
. In
particular as an operator on L
2
the norm M
2
= |K|
of K is bounded by
M
2

_
M
1
M
. (6.7)
The reason for the name interpolation operator (which is not standard) is that
the bound interpolates for all p from the extreme cases p = 1 (where the bound
is M
1
) and p = (where the bound is M
).
Proof: Write
[(Kf)(x)[
_
[k(x, y)[[f(y)[ dy =
_
[k(x, y)[
1
q
[k(x, y)[
1
p
[f(y)[ dy. (6.8)
6.2. BOUNDED LINEAR OPERATORS 83
Apply the Holder inequality to the integral. This gives
[(Kf)(x)[
__
[k(x, y)[ dy
_1
q
__
[k(x, y)[[f(y)[
p
dy
_1
p
M
1
q
__
[k(x, y)[[f(y)[
p
dy
_1
p
.
(6.9)
It follows that
_
[(Kf)(x)[
p
dx M
p
q
_ _
[k(x, y)[[f(y)[
p
dy dx M
p
q
M
1
_
[f(y)[
p
dy.
(6.10)
Take the pth root to get the bound
__
[(Kf)(x)[
p
dx
_1
p
M
1
q
M
1
p
1
__
[f(y)[
p
dy
_1
p
. (6.11)
For a bounded operator K the set of points such that (I K)
1
is
a bounded operator is called the resolvent set of K. The complement of the
resolvent set is called the spectrum of K. When one is only interested in values
of ,= 0 it is common to dene = 1/ and write
(I K)
1
= (I K)
1
(6.12)
While one must be alert to which convention is being employed, this should be
recognized as a trivial relabeling.
Theorem. If complex number satises |K|
< [[, then is in the

resolvent set of K.
Proof: This is the Neumann series expansion
(I K)
1
=
1
j=0
1
j
K
j
. (6.13)
The jth term has norm |K
j
|
/[[
j
(|K|
/[[)
j
. This is the jth term of a
convergent geometric series.
Theorem. If for some n 1 the complex number satises |K
n
|
< [[
n
,
then is in the resolvent set of K.
This is the Neumann series again. But now we only require the estimate
for the nth power of the operator. Write j = an + b, where 0 b < n. Then
|K
j
|
/[[
j
is bounded by (|K|
/[[)
b
((|K
n
|
/[[
n
)
a
. So this is the sum of
n convergent geometric series, one for each value of b between 0 and n 1.
The spectral radius of an operator is the largest value of [[, where is in the
spectrum. The estimate of this theorem gives an upper bound on the spectral
radius.
This is a remarkable result, since it has no analog for scalars. However for a
matrix the norm of a power can be considerably smaller than the corresponding
power of the norm, so this result can be quite useful. The most spectacular
application is to Volterra integral operators.
Example: Volterra integral operators. Let H = L
2
(0, 1) and
(Kf)(x) =
_
x
0
k(x, y)f(y) dy. (6.14)
Suppose that [k(x, y)[ C for 0 y x 1 and k(x, y) = 0 for 0 x < y 1.
Then
|K
n
|

C
n
(n 1)!
. (6.15)
As a consequence every complex number ,= 0 is in the resolvent set of K.
Proof: Each power K
n
with n 1 is also a Volterra integral operator. We
claim that it has kernel k
n
(x, y) with
[k
n
(x, y)[
C
n
(n 1)!
(x y)
n1
. (6.16)
for 0 y x 1 and zero otherwise. This follows by induction.
Since [k
n
(x, y)[ C
n
/(n 1)!, the Hilbert-Schmidt norm of K is also
bounded by C
n
/(n1)!. However this goes to zero faster than every power. So
the Neumann series converges.
The norm of a bounded operator can always be found by calculating the norm
of a self-adjoint bounded operator. In fact, |K|
2
= |K
K|
. Furthermore,
the norm of a self-adjoint operator is its spectral radius. This shows that to
calculate the norm of a bounded operator K exactly, one needs only to calculate
the spectral radius of K
K and take the square root. Unfortunately, this can

be a dicult problem. This is why it is good to have other ways of estimating
the norm.
6.3 Compact operators
Let H be a Hilbert space. A subset S is totally bounded if for every > 0
there is a cover of S by a nite number N of balls. A totally bounded set is
bounded.
Thus for instance, if H is nite dimensional with dimension n and S is a cube
of side L, then N (L/)
n
. This number N is nite, though it increases with
. One expects a cube or a ball to be totally bounded only in nite dimensional
situations.
However the situation is dierent for a rectangular shaped region in innite
dimensions. Say that the sides are L
1
, L
2
, L
3
, . . . , L
k
, . . . and that these decrease
to zero. Consider > 0. Pick k so large that L
k+1
, L
k+1
, . . . are all less than .
Then N (L
1
/)(L
2
/) (L
k
/). This can increase very rapidly with . But
such a region that is fat only in nitely many dimensions and increasingly thin
in all the others is totally bounded.
A linear transformation K : H H is said to be compact if it maps bounded
sets into totally bounded sets. (One can take the bounded set to be the unit
ball.) A compact operator is bounded.
6.3. COMPACT OPERATORS 85
If K and L are compact operators, then so is their sum K + L. If K is
compact and L is bounded, then KL and LK are compact.
If the self-adjoint operator K
K is compact, then K is compact.

Proof: Let > 0. Then there are u
1
, . . . , u
k
in the unit ball such that for
each u in the unit ball there is a j with |K
K(u u
j
)| < /2. But this says
that
|K(uu
j
)| = (K(uu
j
), K(uu
j
)) = ((uu
j
), K
K(uu
j
)) |uu
j
| |K
K(uu
j
)| < .
(6.17)
The adjoint K
of a compact operator K is a compact operator.

Proof: Say that K is compact. Then since K
is bounded, it follows that

KK
is compact. It follows from the last result that K
is compact.
Approximation theorem. If K
n
is a sequence of compact operators, and K
is a bounded operator, and if |K
n
K|
0 and n , then K is also

compact.
The proof is a classical /3 argument. Let > 0. Choose n so large that
|KK
n
|
<

3
. Since K
n
is compact, there are nitely many vectors u
1
, . . . , u
k
in the unit ball such that every vector K
n
u with u in the unit ball is within /3 of
some K
n
u
j
. Consider an arbitrary u in the unit ball and pick the corresponding
u
j
. Since
Ku Ku
j
= (K K
n
)u + (K
n
u K
n
u
j
) + (K
n
K)u
j
, (6.18)
it follows that
|KuKu
j
| |(KK
n
)u|+|K
n
uK
n
u
j
|+|(K
n
K)u
j
|

3
+

3
+

3
= .
(6.19)
This shows that the image of the unit ball under K is totally bounded.
Spectral properties of compact operators. Let K be a compact operator. The
only non-zero points in the spectrum of K are eigenvalues of nite multiplicity.
The only possible accumulation point of the spectrum is 0.
Notice that there is no general claim that the eigenvectors of K form a basis
for the Hilbert space. The example of a Volterra integral operator provides a
counterexample: The only point in the spectrum is zero.
Let K be a compact operator. We want to consider equations of the form
u = f +Ku, (6.20)
where is a parameter. If = 0, then this is an equation of the rst kind. If
,= 0, then this is an equation of the second kind. Very often an equation of
the second kind is written u = f
1
+Ku, where = 1/, and where f
1
= f.
The condition for a unique solution of an equation of the rst kind is the
existence of the inverse K
1
, and the solution is u = K
1
f. Thus the issue
for an equation of the rst kind is whether = 0 is not an eigenvalue of K. If
= 0 is not an eigenvalue, the operator K
1
will be typically be an unbounded
operator that is only dened on a linear subspace of the Hilbert space.
The condition for a unique solution of an equation of the second kind is the
existence of the inverse (I K)
1
for a particular value of ,= 0, and the
solution is u = (I K)
1
f. The issue for an equation of the second kind
is whether ,= 0 is not an eigenvalue of K. In this case, if ,= 0 is not an
eigenvalue, then (I K)
1
will be a bounded operator. Thus equations of
the second kind are much nicer. This is because compact operators have much
better spectral properties away from zero.
Spectral theorem for compact self-adjoint operators. Let K be a compact
self-adjoint operator. Then there is an orthonormal basis u
j
of eigenvectors
of K. The eigenvalues
j
of K are real. Each non-zero eigenvalue is of nite
multiplicity. The only possible accumulation point of the eigenvalues is zero.
The operator K has the representation
Kf =
j
u
j
(u
j
, f). (6.21)
In abbreviated form this is
K =
j
u
j
u
j
. (6.22)
There is yet another way of writing the spectral theorem for compact op-
erators. Dene the unitary operator U from H to
2
by (Uf)
j
= (u
j
, f). Its
inverse is the unitary operator form
2
to H given by (U
c) =

c
j
u
j
. Let M
be the diagonal operator from
2
to
2
dened by multiplication by
j
. Then
K = U
MU. (6.23)
The norm of a compact operator can always be found by calculating the norm
of a self-adjoint compact operator. In fact, |K|
2
= |K
K|
. Furthermore,
the norm of a compact self-adjoint operator is its spectral radius, which in this
case is the largest value of [[, where is an eigenvalue. Unfortunately, this can
be a dicult computation.
Singular value decomposition for compact operators. Let K be a compact
operator. Then there is an orthonormal family u
j
and an orthonormal family
w
j
and a sequence of numbers
j
0 (singular values of K) approaching zero
such that the operator K has the representation
Kf =
j
w
j
(u
j
, f). (6.24)
In abbreviated form this is
K =
j
w
j
u
j
. (6.25)
Sketch of proof: The operator K
K is self-adjoint with positive eigenvalues
2
j
. We can write
K
Kf =
2
j
u
j
(u
j
, f). (6.26)
6.4. HILBERT-SCHMIDT OPERATORS 87
Then

K
Kf =
j
u
j
(u
j
, f). (6.27)
Since |Kf| = |
Kf| for each f, we can write K = V
K, where |V g| =
|g| for all g =

K
Kf in the range of

K
K. This is the well-known polar

decomposition. Then
Kf =
j
w
j
(u
j
, f), (6.28)
where w
j
= V u
j
.
There is another way of writing the singular value decomposition of K. Let
K = U
DU be the spectral representation of

K
K, where D is diagonal
with entries
j
0. Then K = V U
DU = WDU.
It follows from the approximation theorem and from the singular value de-
composition that an operator is compact if and only if it is a norm limit of a
sequence of nite rank operators.
Notice that this theorem gives a fairly clear picture of what a compact op-
erator acting on L
2
looks like. It is an integral operator with kernel
k(x, y) =
j
w
j
(x)u
j
(y), (6.29)
where the
j
0. Of course this representation may be dicult to nd in
practice. What happens in the case of a Hilbert-Schmidt integral operator is
special: the
j
go to zero suciently rapidly that

j
[
j
[
2
< .
6.4 Hilbert-Schmidt operators
Let H be a Hilbert space. For a positive bounded self-adjoint operator B the
trace is dened by trB =

j
(e
j
, Be
j
), where the e
j
form an orthonormal ba-
sis. A bounded linear operator K : H H is said to be Hilbert-Schmidt if
tr(K
K) < . The Hilbert-Schmidt norm is |K|

2
=
_
tr(K
K).
Theorem. If K is a Hilbert-Schmidt integral operator with (Kf)(x) =
_
k(x, y)f(y) dy then
|K|
2
2
=
_ _
[k(x, y)[
2
dxdy. (6.30)
Proof: Let e
1
, e
2
, ... be an orthonormal basis for H. We can always expand
a vector in this basis via
u =
j
e
j
(e
j
, u). (6.31)
(Remember the convention adopted here that inner products are linear in the
second variable, conjugate linear in the rst variable.) Write
Kf =
i
e
i
j
(e
i
, Ke
j
)(e
j
, f). (6.32)
The matrix elements (e
i
, Ke
j
) satisfy
j
[(e
i
, Ke
j
)[
2
=
i
(Ke
j
, e
i
)(e
i
, Ke
j
) =
j
[[Ke
j
[[
2
=
j
(e
j
, K
Ke
j
) = tr(K
K).
(6.33)
So the kernel of the integral operator K is
k(x, y) =
j
(e
i
, Ke
j
)e
i
(x)e
j
(y). (6.34)
This sum is convergent in L
2
(dxdy).
A Hilbert-Schmidt operator is a bounded operator. It is always true that
|K|
|K|
2
.
Theorem. A Hilbert-Schmidt operator is compact.
Proof: Let K be a Hilbert-Schmidt operator. Then K is given by a square-
summable matrix. So K may be approximated by a sequence of nite-rank
operators K
n
such that |K
n
K|
2
0. In particular, |K
n
K|
0.
Since each K
n
is compact, it follows from the approximation theorem that K is
compact.
If K and L are Hilbert-Schmidt operators, their sum K + L is a Hilbert-
Schmidt operator and |K +L|
2
|K|
2
+|L|
2
. If K is a Hilbert-Schmidt op-
erator and L is a bounded operator, then the products KL and LK are Hilbert-
Schmidt. Furthermore, |KL|
2
|K|
2
|L|
and |LK|
2
|L|
|K|
2
.
The adjoint of a Hilbert-Schmidt operator is a Hilbert-Schmidt operator.
Furthermore, |K
|
2
= |K|
2
.
Notice that the Hilbert-Schmidt norm of a bounded operator is dened in
terms of a self-adjoint operator. In fact, |K|
2
is the square root of the trace
of the self-adjoint operator K
K, and the trace is the sum of the eigenvalues.

However we do not need to calculate the eigenvalues, since, as we have seen,
there are much easier ways to calculate the trace.
6.5 Problems
1. Let H = L
2
be the Hilbert space of square integrable functions on the
line. Create an example of a Hilbert-Schmidt operator that is not an
interpolation operator.
2. Let H = L
2
be the Hilbert space of square integrable functions on the
line. Create an example of an interpolation operator that is not a Hilbert-
Schmidt operator. Make the example so that the operator is not compact.
3. Find an example of a compact bounded operator on H = L
2
that is neither
a Hilbert-Schmidt or an interpolation operator.
4. Find an example of a bounded operator on H = L
2
that is neither a
Hilbert-Schmidt nor an interpolation operator, and that is also not a com-
pact operator.
6.6. FINITE RANK OPERATORS 89
5. Let H be a Hilbert space. Give an example of a Hilbert-Schmidt operator
for which the spectral radius is equal to the uniform norm.
6. Let H be a Hilbert space. Give an example of a Hilbert-Schmidt operator
for which the spectral radius is very dierent from the uniform norm.
7. Let H be a Hilbert space. Is it possible for a Hilbert-Schmidt operator
to have its Hilbert-Schmidt norm equal to its uniform norm? Describe all
possible such situations.
6.6 Finite rank operators
Let H be a Hilbert space. A bounded linear transformation K : H H is said
to be nite rank if its range is nite dimensional. The dimension of the range
of K is called the rank of K. A nite rank operator may be represented in the
form
Kf =
j
z
j
(u
j
, f), (6.35)
where the sum is nite. In a more abbreviated notation we could write
K =
j
z
j
u
j
. (6.36)
Thus in L
2
this is an integral operator with kernel
k(x, y) =
j
z
j
(x)u
j
(y). (6.37)
If K and L are nite rank operators, then so is their sum K + L. If K is
nite rank and L is bounded, then KL and LK are nite rank.
The adjoint K
of a nite rank operator K is nite rank. The two operators

have the same rank.
Every nite rank operator is Hilbert-Schmidt and hence compact. If K is a
compact operator, then there exists a sequence of nite rank operators K
n
such
that |K
n
K|
0 as n .
If K is a nite rank operator and ,= 0, then the calculation of (I K)
1
or of (I K)
1
may be reduced to a nite-dimensional matrix problem in
the nite dimensional space R(K). This is because K leaves R(K) invariant.
Therefore if = 1/ is not an eigenvalue of K acting in R(K), then there is an
inverse (I K)
1
acting in R(K). However this gives a corresponding inverse
in the original Hilbert space, by the formula
(I K)
1
= I + (I K)
1
K. (6.38)
Explictly, to solve
u = Ku +f, (6.39)
write
u = (I K)
1
f = f + (I K)
1
Kf. (6.40)
To solve this, let w be the second term on the right hand side, so that u = f +w.
Then (I K)w = Kf. Write w =

j
a
j
z
j
. Then
j
a
j
z
j

j
z
j
(u
j
,
r
a
r
z
r
) =
j
z
j
(u
j
, f). (6.41)
Thus
a
j

r
(u
j
, z
r
)a
r
= (u
j
, f). (6.42)
This is a matrix equation that may be solved whenever 1/ is not an eigenvalue
of the matrix with entries (u
j
, z
r
).
6.7 Problems
It may help to recall that the problem of inverting I K is the same as the
problem of showing that = 1/ is not in the spectrum of K.
1. Consider functions in L
2
(, ). Consider the integral equation
f(x)
_

cos(
_
x
2
+y
4
)e
|x||y|
f(y) dy = g(x).
It is claimed that there exists r > 0 such that for every complex number
with [[ < r the equation has a unique solution. Prove or disprove. Inter-
pret this as a statement about the spectrum of a certain linear operator.
2
f(x)
_

cos(
_
x
2
+y
4
)e
|x||y|
f(y) dy = g(x).
It is claimed that there exists R < such that for every complex number
with [[ > R the equation does not have a unique solution. Prove or
disprove. Interpret this as a statement about the spectrum of a certain
linear operator.
2
f(x)
_

e
|x||y|
f(y) dy = g(x).
Find all complex numbers for which this equation has a unique solution.
Find the solution. Interpret this as a statement about the spectrum of a
certain linear operator.
6.7. PROBLEMS 91
2
(0, 1). Consider the integral equation
f(x)
_
x
0
f(y) dy = g(x).
Find the solution. Interpret this as a statement about the spectrum of a
certain linear operator. Hint: Dierentiate. Solve a rst order equation
with a boundary condition.
2
(0, 1). Consider the integral equation
f(x)
__
x
0
y(1 x)f(y) dy +
_
1
x
x(1 y)f(y) dy
_
= g(x).
Interpret this as a statement about the spectrum of a certain linear oper-
ator. Hint: The integral operator K has eigenfunctions sin(nx). Verify
this directly. This should also determine the eigenvalues.
Chapter 7
Densely Dened Closed
Operators
7.1 Introduction
This chapter deals primarily with densely dened closed operators. Each every-
where dened bounded operator is in particular a densely dened closed oper-
ator.
If L is a densely dened closed operator, then so is its adjoint L
. Further-
more, L
= L. If both L and L
have trivial null spaces, then both L

1
and
L
1
are densely dened closed operators.
A complex number is in the resolvent set of a densely dened closed op-
erator L if (L I)
1
is an everywhere dened bounded operator. A complex
number is in the spectrum of L if it is not in the resolvent set. A complex
number is in the spectrum of L if and only if

is in the spectrum of the
adjoint L
.
It is common to divide the spectrum into three disjoint subsets: point spec-
trum, continuous spectrum, and residual spectrum. (This terminology is mis-
leading, in that it treats limits of point spectrum as continuous spectrum.) In
this treatment we divide the spectrum into four disjoint subsets: standard point
spectrum, pseudo-continuous spectrum, anomalous point spectrum, and resid-
ual spectrum. The adjoint operation maps the rst two kind of spectra into
themselves, but it reverses the latter two.
7.2 Subspaces
Let H be a Hilbert space. Let M be a vector subspace of H. The closure

M
is also a vector subspace of H. The subspace M is said to be closed if M =
M. The orthogonal complement M
is a closed subspace of H. Furthermore,

(

M)
= M
. Finally M
=

M. The nicest subspaces are closed subspaces.
93
94 CHAPTER 7. DENSELY DEFINED CLOSED OPERATORS
For a closed subspace M we always have M
= M.
A subspace M is dense in H if

M = H. This is equivalent to the condition
M
= 0.
7.3 Graphs
Let H be a Hilbert space. The direct sum H H with itself consists of all
ordered pairs [u, v], where u, v are each vectors in H. The inner product of
[u, v] with [u
, v
] is
([u, v], [u
, v
]) = (u, u
) + (v, v
). (7.1)
A graph is a linear subspace of H H. If we have two graphs L
1
and L
2
and if L
1
L
2
, then we say that L
1
is a restriction of L
2
or L
2
is an extension
of L
1
.
If L is a graph, then its domain D(L) is the set of all u in H such that there
exists a v with [u, v] in L. Its range R(L) is the set of all v in H such that there
exists u with [u, v] in L.
If L is a graph, then its null space N(L) is the set of all u in H such that
[u, 0] is in L.
If L is a graph, then the inverse graph L
1
consists of all [v, u] such that
[u, v] is in L.
We say that a graph L is an operator if [0, v] in L implies v = 0. It is easy
to see that this is equivalent to saying that N(L
1
) = 0 is the zero subspace.
When L is an operator and [u, v] is in L, then we write Lu = v. We shall explore
the properties of operators in the next section.
Write

L for the closure of L. We say L is closed if L =

L.
If L is a graph, then the adjoint graph L
consists of the pairs [w, z] such

that for all [u, v] in L we have (z, u) = (w, v).
If L is a graph, then the adjoint graph L
is always closed. Furthermore,

the adjoint of its closure

L is the same as the adjoint of L.
Remark: One way to think of the adjoint graph is to dene the negative
inverse L
1
of a graph L to consist of all the ordered pairs [v, u] with [u, v]
in L. Then the adjoint L
is the orthogonal complement in H H of L

1
.
That is, the pair [z, w] in the graph L
is orthogonal to each [v, u] with [u, v]

in the graph of L. This says that ([z, w], [v, u]) = (z, u) (w, v) = 0 for all
such [u, v]. [This says to take the graph with negative reciprocal slope, and then
take the perpendicular graph to that.]
Another way to think of this is to dene the anti-symmetric form([z, w], [u, v]) =
(z, u) (w, v). Then the adjoint A
consists of the orthogonal complement of

A with respect to .
Theorem. If L is a graph, then L
=

L.
Perhaps the nicest general class of graphs consists of the closed graphs L.
For such a graph the adjoint L
is a graph, and L
= L.
It is not hard to check that L
1
= L
1
.
Theorem. N(L
) = R(L)
.
7.4. OPERATORS 95
Corollary. R(L) = N(L
.
This corollary is very important in the theory of linear equations. Let L be
a linear operator. Suppose that R(L) is a closed subspace of H. Then in this
special case the corollary says that R(L) = N(L
. Thus a linear equation

Lu = g has a solution u if and only if g is orthogonal to all solutions v of
L
v = 0. This is called the Fredholm alternative.

7.4 Operators
If L is a graph, then L is an operator if [0, v] in L implies v = 0. It is easy to
see that L is an operator precisely when N(L
1
) = 0 is the zero subspace.
When L is an operator and [u, v] is in L, then we write Lu = v.
Corollary. L is densely dened if and only if L
is an operator.
Proof: Apply the last theorem of the previous section to L
1
. This gives
N(L
1
) = D(L)
.
Corollary.

L is an operator if and only if L
is densely dened.
Proof: Apply the previous corollary to L
and use L
=

L.
An operator L is said to be closable if

L is also an operator. For a densely
dened closable operator the adjoint L
is a densely dened closed operator,

and L
=

L. Furthermore, the denition of the adjoint is that w is in the
domain of L
and L
w = z if and only if for all u in the domain of L we have

(z, u) = (w, Lu), that is,
(L
w, u) = (w, Lu). (7.2)

Perhaps the nicest general class of operators consists of the densely dened
closed operators L. For such an operator the adjoint L
is a densely dened
closed operator, and L
= L.
7.5 The spectrum
Theorem. (Closed graph theorem) Let H be a Hilbert space. Let L be a closed
operator with domain D(L) = H. Then L is a bounded operator. (The converse
is obvious.)
0. Let L be a closed operator. We say that is in the resolvent set of L if
N(LI) = 0 and R(LI) = H. In that case (LI)
1
is a closed operator
with domain D((L I)
1
) = H. By the closed graph theorem, (L I)
1
is
a bounded operator.
We shall usually refer to (LI)
1
as the resolvent of L. However in some
contexts it is convenient to use instead (I L)
1
, which is of course just the
negative. There is no great distinction between these two possible denitions of
resolvent. However it is important to be alert to which one is being used.
1. Let L be a closed operator. We say that is in the standard point
spectrum of L if N(L I) ,= 0 and R(L I)
,= 0.
2. Let L be a closed operator. We say that is in the anomalous point
spectrum of L if N(L I) ,= 0 and R(L I)
= 0 (that is, R(L I)

is dense in H).
3. Let L be a closed operator. We say that is in the pseudo-continuous
spectrum of L if N(LI) = 0 and R(LI)
= 0 (so R(LI) is dense

in H) but R(L I) ,= H. In that case (L I)
1
is a closed operator with
dense domain D((L I)
1
) not equal to H.
4. Let L be a closed operator. We say that is in the residual spectrum
of L if N(L I) = 0 and R(L I)
,= 0. In that case (L I)
1
is a
closed operator with a domain that is not dense.
Theorem. Let L be a densely dened closed operator and let L
be its adjoint
operator. Then:
0. The number is in the resolvent set of L if and only if

is in the resolvent
set of L
.
1. The number is in the standard point spectrum of L if and only if

is
in the standard point spectrum of L
.
2. The number is in the anomalous point spectrum of L if and only if

is
in the residual spectrum of L
.
3. is in the pseudo-continuous spectrum of L if and only if

is in the
pseudo-continuous spectrum of L
.
4. is in the residual spectrum of L if and only if

is in the anomalous
point spectrum of L
.
For nite dimensional vector spaces only cases 0 and 1 can occur.
Summary: Let L be a closed, densely dened operator. The complex number
is in the point spectrum of L is equivalent to being an eigenvalue of L.
Similarly, in the pseudo-continuous spectrum of L is equivalent to (LI)
1
being a densely dened, closed, but unbounded operator. Finally, in the
residual spectrum of L is equivalent to (LI)
1
being a closed operator that
is not densely dened.
7.6 Spectra of inverse operators
Consider a closed, densely dened operator with an inverse L
1
that is also a
closed, densely dened operator. Let ,= 0. Then is in the resolvent set of L
if and only if 1/ is in the resolvent set of L
1
. In fact, we have the identity
_
I
1
L
_
1
+
_
I L
1
_
1
= I. (7.3)
One very important situation is when K = L
1
is a compact operator.
Then we know that all non-zero elements of the spectrum of K = L
1
are
eigenvalues of nite multiplicity, with zero as their only possible accumulation
point. It follows that all elements = 1/ of the spectrum of L are eigenvalues
of nite multiplicity, with innity as their only possible accumulation point.
7.7. PROBLEMS 97
7.7 Problems
If K is a bounded everywhere dened operator, then in particular K is a closed
densely dened operator.
If K is a closed densely dened operator, and if both K and K
have trivial
nullspaces, then L = K
1
is also a closed densely dened operator.
1. Let K = L
1
be as above. Let ,= 0 and let = 1/. Find a formula
relating the resolvent (L )
1
to the resolvent (K )
1
.
2
(0, 1). Consider the integral operator K given by
(Kf)(x) =
_
x
0
f(y) dy.
Show that L = K
1
exists and is closed and densely dened. Describe
the domain of L. Be explicit about boundary conditions. Describe how L
acts on the elements of this domain. Show that L
= K
1
is closed and
densely dened. Describe the domain of L
. Describe how L
acts on the
elements of this domain. Hint: Dierentiate.
3. In the preceding problem, nd the spectrum of L. Also, nd the resol-
vent (L )
1
of L. Hint: Solve a rst order linear ordinary dierential
equation.
2
(0, 1). Consider the integral operator K given by
(Kf)(x) =
__
x
0
y(1 x)f(y) dy +
_
1
x
x(1 y)f(y) dy
_
.
Show that L = K
1
exists and is closed and densely dened. Describe
the domain of L. Be explicit about boundary conditions. Describe how L
acts on the elements of this domain. Hint: Dierentiate twice.
5. In the preceding problem, nd the spectrum of L. Find the resolvent
(L )
1
of L. Hint: Use sin(
x) and sin(
(1 x)) as a basis for

the solutions of a homogeneous second order linear ordinary dierential
equation. Solve the inhomogeneous equation by variation of parameters.
6. Let K be a compact operator. Suppose that K and K
have trivial null-

spaces, so that L = K
1
is a closed densely dened operator. Prove
that the spectrum of L = K
1
consists of isolated eigenvalues of nite
multiplicity. To what extent does this result apply to the examples in the
previous problems?
7. Let K be a compact self-adjoint operator. Suppose that K has trivial
null-space, so that L = K
1
is a self-adjoint operator. Prove that there
exists an orthogonal basis consisting of eigenvectors of L. To what extent
does this result apply to the examples in the previous problems?
7.8 Self-adjoint operators
It is dicult to do algebraic operations with closed, densely dened operators,
because their domains may dier. It is always true that (zL)
= zL
. If K is
bounded everywhere dened, then (L+K)
= L
+K
and (K+L)
= K
+L
.
Furthermore, if K is bounded everywhere dened, then (KL)
= L
.
An operator A is self-adjoint if A = A
. A self-adjoint operator is automat-

ically closed and densely dened. (Every adjoint A
is automatically closed. If
A
is an operator, then A is densely dened.)

If a self-adjoint operator has trivial null space, then its inverse is also a
self-adjoint operator.
If L is a closed, densely dened operator, then L
L is dened on the domain

consisting of all u in D(L) such that Lu is in D(L
). It is not obvious that this

is closed and densely dened, much less that it is self-adjoint. However this is
all a consequence of the following theorem.
Theorem. If L is a closed and densely dened operator, then L
L is a self-
adjoint operator.
Proof: If L is a closed and densely dened operator, then L
is also a closed
and densely dened operator. Furthermore, LL
is an operator with LL

(LL
.
The Hilbert space H H may be written as the direct sum of the two
closed graphs L
1
and L
Therefore an arbitrary [0, h] for h in H may be

written as the sum [0, h] = [Lf, f] + [g, L
g]. This says that 0 = Lf + g

and h = f +L
g. As a consequence h = f +L
Lf. Furthermore, by properties

of projections we have |Lf|
2
+ |f|
2
|h|
2
. We have shown that for each
h we can solve (I + L
L)f = h and that |f|

2
|h|
2
. Thus (I + L
L)
1
is
everywhere dened and is a bounded operator with norm bounded by one.
Since L
L (L
L)
, we have (I + L
L)
1
(I + L
L)
1
. It follows that
(I +L
L)
1
= (I +L
L)
1
is a self-adjoint operator. The conclusion follows.
7.9 First order dierential operators with a bounded
interval: point spectrum
In this section we shall see examples of operators with no spectrum at all.
However we shall also see a very pretty and useful example of an operator with
standard point spectrum. This operator is the one behind the theory of Fourier
series.
Example 1A: This example is one where the correct number of boundary
conditions are imposed. In the case of a rst order dierential operator this
number is one. Let H be the Hilbert space L
2
(0, 1). Let L
0
be the operator
d/dx acting on functions of the form f(x) =
_
x
0
g(y) dy where g is in H. The
value of L
0
on such a function is g(x). Notice that functions in the domain of L
0
automatically satisfy the boundary condition f(0) = 0. This is an example of a
closed operator. The reason is that (L
1
0
g)(x) =
_
x
0
g(y) dy. This is a bounded
operator dened on the entire Hilbert space. So L
1
0
and L
0
are both closed.
7.9. FIRST ORDER DIFFERENTIAL OPERATORS WITHABOUNDEDINTERVAL: POINT SPECTRUM99
The adjoint of the inverse is given by (L
1
0
)h(x) =
_
1
x
h(y) dy. It follows
that L
0
is the operator L
1
, where L
1
is given by d/dx acting on functions
of the form f(x) =
_
1
x
g(y) dy where g is in H. The value of L
1
on such a
function is g(x). Notice that functions in the domain of L
1
automatically satisfy
the boundary condition f(1) = 0.
The operators L
0
and L
1
each have one boundary condition. They are
negative adjoints of each other. They each have a spectral theory, but it is
extremely pathological. For instance, the resolvent of L
0
is given by
((L
0
I)
1
g)(x) =
_
x
0
e
(xy)
g(y) dy.
So there are no points at all in the spectrum of L
0
. It is in some sense located
all at innity.
To see this, consider the operator L
1
0
. This operator has spectrum consist-
ing of the point zero. All the spectral information is hidden at this one point.
This is, by the way, an example of pseudo-continuous spectrum.
This is one important though somewhat technical point. The domain of L
0
consists precisely of the functions in the range of K
0
= L
1
0
. In the example
where K
0
is the integration operator, this is all functions of the form
u(x) =
_
x
0
f(y)dy, (7.4)
where f is in L
2
(0, 1). These functions u need not be C
1
. They belong to a larger
class of functions that are indenite integrals of L
2
functions. Such functions
are continuous, but they may have slope discontinuities. The functions u of
course satisfy the boundary condition u(0) = 0. The action of L
0
on a function
u is given by L
0
u = u
, where the derivative exists except possible on a set of

measure zero. However L
2
functions such as f are dened only up to sets of
measure zero, so this is not a problem.
Now for a really picky question: If u is also regarded as an L
2
function, then
it is also dened only up to sets of measure zero. So what does u(0) mean?
After all, the set consisting of 0 alone is of measure zero. The answer is that
the general indenite integral is a function of the form
u(x) =
_
x
0
f(y)dy +C. (7.5)
Among all L
2
functions given by such an integral expression, there is a subclass
of those for which C = 0. These are the ones satisfying the boundary condition
u(0) = 0. There is another subclass for which C =
_
1
0
f(y)dy. These are the
ones satisfying u(1) = 0.
A densely dened operator L is said to be self-adjoint if L = L
. Similarly,
L is said to be skew-adjoint if L = L
.
Example 1B: Here is another example with the correct number of boundary
conditions. Let H be the Hilbert space L
2
(0, 1). Let L
=
be the operator d/dx
acting on functions of the form f(x) =
_
x
0
g(y) dy + C where g is in H and
_
1
0
g(y) dy = 0. The value of L
=
on such a function is g(x). Notice that
functions in the domain of L
=
automatically satisfy the boundary condition
f(0) = f(1). The operator L
=
is skew-adjoint.
The operator L
=
has periodic boundary conditions. Since it is skew-adjoint,
it has an extraordinarily nice spectral theory. The resolvent is
((L
=
I)
1
g)(x) =
1
1 e
_
x
0
e
(xy)
g(y) dy
1
1 e
_
1
x
e
(yx)
g(y) dy.
The spectrum consists of the numbers 2in. These are all point spectrum. The
corresponding eigenvectors form a basis that gives the Fourier series expansion
of an arbitrary periodic function with period one.
A densely dened operator L is said to be Hermitian if L L
. This is
simply the algebraic property that
(Lu, w) = (u, Lw)
for all u, w in D(L). Similarly, L is said to be skew-Hermitian if L L
.
Example 2: The following example illustrates what goes wrong when one
imposes the wrong number of boundary conditions. Let H be the Hilbert space
L
2
(0, 1). Let L
01
be the operator d/dx acting on functions of the form f(x) =
_
x
0
g(y) dy where g is in H and
_
1
0
g(y) dy = 0. The value of L
01
on such a
function is g(x). Notice that functions in the domain of L
01
automatically
satisfy the boundary conditions f(0) = 0 and f(1) = 0. The adjoint of L
01
is the operator L, where L is given by d/dx acting on functions of the form
_
x
0
g(y) dy+C, where g is in H. The value of L on such a function is g(x). Notice
that functions in the domain of L need not satisfy any boundary conditions.
From this we see that L
01
is skew-Hermitian.
The operator L is d/dx. It has too few boundary conditions. The opera-
tor L
01
has a boundary condition at 0 and at 1. This is too many boundary
conditions. Each of these operators is the negative of the adjoint of the other.
The spectrum of L consists of the entire complex plane, and it is all point spec-
trum. The spectrum of L
01
also consists of the entire complex plane, and it is
all residual spectrum.
Remark: The operators L
0
and L
1
have L
01
L
0
L and with L
01
L
1

L. Furthermore, the operator L
=
has L
01
L
=
L. Thus there are various
correct choices of boundary conditions, but they may have dierent spectral
properties.
7.10 Spectral projection and reduced resolvent
Consider a closed densely dened operator L. In this section we shall assume
that the eigenvectors of L span the entire Hilbert space. Consider also an
isolated eigenvalue
n
. The spectral projection corresponding to
n
is a (not
7.11. GENERATING SECOND-ORDER SELF-ADJOINT OPERATORS101
necessarily orthogonal) projection onto the corresponding eigenspace. It is given
in terms of the resolvent by
P
n
= lim
n
(
n
)(L I)
1
. (7.6)
This is the negative of the residue of the resolvent at
1
. The reason this works
is that L =

m
m
P
m
and consequently
(L I)
1
=
m
1
P
m
, (7.7)
at least in the case under consideration, when the eigenvectors span the entire
Hilbert space.
The reduced resolvent corresponding to
n
is dened as the operator that
inverts (L
n
I) in the range of (I P
n
) and is zero in the range of P
n
. It is
a solution of the equations
S
n
P
n
= P
n
S
n
= 0 (7.8)
and
(L
n
I)S
n
= 1 P
n
(7.9)
It may be expressed in terms of the resolvent by
S
n
= lim
n
(L I)
1
(1 P
n
). (7.10)
When the eigenvectors span this is
S
n
=
m=n
1
n
P
m
. (7.11)
Example: Take the skew-adjoint operator L
=
= d/dx acting in L
2
(0, 1)
with periodic boundary conditions. The spectral projection corresponding to
eigenvalue 2in is the self-adjoint operator
(P
n
g)(x) =
_
1
0
exp(2in(x y))g(y) dy. (7.12)
The reduced resolvent corresponding to eigenvalue 0 is the skew-adjoint operator
(S
0
g)(x) =
_
x
0
(
1
2
x +y)g(y) dy +
_
1
x
(
1
2
x +y)g(y) dy. (7.13)
7.11 Generating second-order self-adjoint oper-
ators
This section exploits the theorem that says that if L is an arbitrary closed
densely dened operator, then L
L is a self-adjoint operator. Remember that

L
= L, so LL
is also a self-adjoint operator.

It would be easy to conclude that rst order dierential operators such as L
01
and its adjoint L are of no interest for spectral theory. This is not the case. From
the general theorem LL
01
and L
01
L are self-adjoint second-order dierential
operators. These, as we shall see, have a nice spectral theory. The operator
LL
01
is the operator d
2
/dx
2
with Dirichlet boundary conditions u(0) = 0 and
u(1) = 0. The operator L
01
L is the operator d
2
/dx
2
with Neumann boundary
conditions u
(0) = 0 and u
(1) = 0. It is amusing to work out other self-adjoint

second order dierential operators that may be generated from the rst order
dierential operators of the preceding sections.
7.12 First order dierential operators with a semi-
innite interval: residual spectrum
In this example we shall see examples of operators with anomalous point spec-
trum and residual spectrum. These operators underly the theory of the Laplace
transform.
Example: Let H be the Hilbert space L
2
(0, ). Let L
0
be the operator d/dx
acting on functions f in H of the form f(x) =
_
x
0
g(y) dy where g is in H. The
value of L
0
on such a function is g(x). Notice that functions in the domain of
L
0
automatically satisfy the boundary condition f(0) = 0.
If ' < 0, then we can nd always nd a solution of the equation (L
0

I)f = g. This solution is
f(x) =
_
x
0
e
(xy)
g(y) dy. (7.14)
This equation denes the bounded operator (L
0
I)
1
that sends g into f, at
least when ' < 0. Notice that if ' > 0, then the formula gives a result in L
2
only if g is orthogonal to e
x
. Thus ' > 0 corresponds to residual spectrum.
Again let H be the Hilbert space L
2
(0, ). Let L be the operator d/dx with
no boundary condition. If ' > 0, then we can nd always nd a solution of
the equation (L I)f = g. This solution is
f(x) =
_

x
e
(xy)
g(y) dy. (7.15)
This equation denes the bounded operator (L I)
1
least when ' > 0. On the other hand, if ' < 0, then we have point spectrum.
The relation between these two operators is L
0
= L. This corresponds to
the fact that (L
0
I)
1
= (L
)
1
, which is easy to check directly.
7.13 First order dierential operators with an
innite interval: continuous spectrum
In this section we shall see an example of an operator with continuous spectrum.
This is the example that underlies the theory of the Fourier transform.
7.14. PROBLEMS 103
Example: Let H be the Hilbert space L
2
(, ). Let L be the operator
d/dx acting on functions f in H with derivatives f
in H.
If ' < 0, then we can nd always nd a solution of the equation (LI)f =
g. This solution is
f(x) =
_
x
e
(xy)
g(y) dy. (7.16)
1
least when ' < 0. If ' > 0, then we can nd always nd a solution of the
equation (L I)f = g. This solution is
f(x) =
_

x
e
(xy)
g(y) dy. (7.17)
1
least when ' > 0.
The operator L is skew-adjoint, that is, L
= L. This corresponds to the

fact that (L I)
1
= (L
)
1
, which is easy to check directly.
7.14 Problems
1. Let H = L
2
(0, 1). Let L = id/dx with periodic boundary conditions.
Find an explicit formula for (L)
1
g. Hint: Solve the rst order ordinary
dierential equation ( L)f = g with the boundary condition f(0) =
f(1).
2. Find the eigenvalues and eigenvectors of L. For each eigenvalue
n
, nd
the residue P
n
of ( L)
1
at
n
.
3. Find the explicit form of the formula g =

n
P
n
g.
4. Let H = L
2
(, ). Let L = id/dx. Let k be real and > 0. Find
an explicit formula for (L k i)
1
g. Also, nd an explicit formula for
(L k +i)
1
g. Find the explicit form of the expression
(L k)g =
1
2i
[(L k i)
1
(L k +i)
1
]g.
5. Find the explicit form of the formula
g =
_

(L k)g dk.
6. Let 0. Find the explicit form of the formula
g =
_

(L k)g dk.
7.15 A pathological example
Consider H = L
2
(R) and x g in H. Dene K as the integral operator
with kernel k(x, y) = g(x)(y). Consider the domain D(K) to be the set
of all continuous functions in H. Then
(Ku)(x) = g(x)f(0). (7.18)
This K is densely dened but not closed. It has a closure

K, but this is
not an operator. To see this, let u
n
u and Ku
n
= u
n
(0)g v. Then
the pair [u, v] is in the graph

K. But we can take u
n
0 in the Hilbert
space sense, yet with each u
n
(0) = C. So this gives the pair [0, Cg] in the
graph

K. This is certainly not an operator!
Consider the adjoint K
. This is the integral operator with kernel (x)g(y).

That is,
(K
w)(x) = (x)
_

g(y)w(y) dy. (7.19)

Since
_

(x)
2
dx = +, (7.20)
the (x) is not in L
2
. Hence the domain of K
consists of all w with

(g, w) = 0, and K
= 0 on this domain. This is an operator that is closed

but not densely dened. According to the general theory, its adjoint is
K
=

K, which is not an operator.
Chapter 8
Normal operators
8.1 Spectrum of a normal operator
Theorem (von Neumann) Let L be a densely dened closed operator. Then L
L
and LL
are each self-adjoint operators.

A densely dened closed operator is said to be normal if L
L = LL
.
There are three particularly important classes of normal operators.
1. A self-adjoint operator is an operator L with L
= L.
2. A skew-adjoint operator is an operator L with L
= L.
3. A unitary operator is an operator L with L
= L
1
. A unitary operator
is bounded.
For a self-adjoint operator the spectrum is on the real axis. For a skew-
adjoint operator the spectrum is on the imaginary axis. For a unitary operator
the spectrum is on the unit circle.
For normal operators there is a dierent classication of spectrum. Let L be
a normal operator acting in a Hilbert space H. The point spectrum consists of
the eigenvalues of L. The corresponding eigenvectors span a closed subspace M
p
of the Hilbert space. The spectrum of L in this space consists of either what we
have previously called standard point spectrum or of what we have previously
called pseudo-continuous spectrum. This kind of pseudo-continuous spectrum
is not really continuous at all, since it consists of limits of point spectrum.
Let M
c
be the orthogonal complement in H of M
p
. Then the spectrum
of L restricted to M
c
is called the continuous spectrum of L. In our previ-
ous classication the spectrum of L in this space would be pseudo-continuous
spectrum.
With this classication for normal operators the point spectrum and contin-
uous spectrum can overlap. But they really have nothing to do with each other,
since they take place in orthogonal subspaces.
Spectral theorem for compact normal operators. Let K be a compact normal
operator. (This includes the cases of self-adjoint operators and skew-adjoint
operators.) Then K has an orthogonal basis of eigenvectors. The non-zero
105
106 CHAPTER 8. NORMAL OPERATORS
eigenvalues have nite multiplicity. The only possible accumulation point of
eigenvalues is zero.
Spectral theorem for normal operators with compact resolvent. Let L be
a normal operator with compact resolvent. (This includes the cases of self-
adjoint operators and skew-adjoint operators.) Then L has an orthogonal basis
of eigenvectors. The eigenvalues have nite multiplicity. The only possible
accumulation point of eigenvalues is innity.
8.2 Problems
1. Perhaps the most beautiful self-adjoint operator is the spherical Laplacian
S
=
1
sin
sin

+
1
sin
2
2
.
Show by explicit computation that this is a Hermitian operator acting on
L
2
of the sphere with surface measure sin d d. Pay explicit attention
to what happens at the north pole and south pole when one integrates by
parts.
2. Let r be the radius satisfying r
2
= x
2
+y
2
+z
2
. Let
L = r

r
be the Euler operator. Show that the Laplace operator
=

2
x
2
+

2
y
2
+

2
z
2
is related to L and
S
by
=
1
r
2
[L(L + 1) +
S
].
3. Let p be a polynomial in x, y, z that is harmonic and homogeneous of
degree . Thus p = 0 and Lp = p. Such a p is called a solid spherical
harmonic. Show that each solid spherical harmonic is an eigenfunction of
S
and nd the corresponding eigenvalue as a function of .
4. The restriction of a solid spherical harmonic to the sphere r
2
= 1 is called
a surface spherical harmonic. The surface spherical harmonics are the
eigenfunctions of
S
. Show that surface spherical harmonics for dierent
values of are orthogonal in the Hilbert space of L
2
functions on the
sphere.
5. The dimension of the eigenspace indexed by is 2 + 1. For = 0 the
eigenspace is spanned by 1. For = 1 it is spanned by z, x + iy, and
x iy. For = 2 it is spanned by 3z
2
r
2
, z(x +iy), z(x iy), (x +iy)
2
,
8.3. VARIATION OF PARAMETERS AND GREENS FUNCTIONS 107
and (x iy)
2
. For = 3 it is spanned by 5z
3
3zr
2
, (5z
2
r
2
)(x + iy).
(5z
2
r
2
)(x iy), z(x +iy)
2
, z(x iy)
2
, (x +iy)
3
, (x iy)
3
. Express the
corresponding surface spherical harmonics in spherical coordinates.
6. In the case = 1 we can write the general spherical harmonic as ax+by+cz.
In the case = 2 we can write it as ax
2
+ by
2
+ cz
2
+ dxy + eyz + fzx
with an additional condition on the coecients. What is this condition?
In the case = 3 we can write it as a
1
x
3
+b
1
y
2
x+c
1
z
2
x+a
2
y
3
+b
2
z
2
y +
c
2
x
2
y +a
3
z
3
+b
3
x
2
z +c
3
y
2
z +dxyz with additional conditions. What are
they?
8.3 Variation of parameters and Greens func-
tions
First look at rst order linear ordinary dierential operators. Let Lu = p(x)u
+
r(x)u. Let u
1
be a non-zero solution of the homogeneous equation Lu = 0. The
general solution of the homogeneous equation Lu = 0 is u(x) = c
1
u
1
(x), where
c
1
is a parameter. The method of variation of parameters gives a solution of
the inhomogeneous equation Lu = f in the form u(x) = c
1
(x)u
1
(x).
The condition on the parameter is given by plugging u into Lu = f. This
gives p(x)c
1
(x)u
1
(x) = f(x). The solution is c
1
(x) = f(x)/(p(x)u
1
(x)). The
only dicult part is to integrate this to get the general solution
u(x) =
_
x
a
u
1
(x)
p(y)u
1
(y)
f(y) dy +Cu
1
(x). (8.1)
Now look at second order linear ordinary dierential operators. Let Lu =
p(x)u
+r(x)u
+q(x)u. Let u
1
and u
2
be independent solutions of the homoge-
neous equation Lu = 0. The general solution of the homogeneous equation Lu =
0 is u(x) = c
1
u
1
(x) +c
2
u
2
(x), where c
1
and c
2
are parameters. The method of
variation of parameters gives a solution of the inhomogeneous equation Lu = f
in the form u(x) = c
1
(x)u
1
(x) +c
2
(x)u
2
(x). Not only that, it has the property
that the derivative has the same form, that is, u
(x) = c
1
(x)u
1
(x) +c
2
(x)u
2
(x).
If this is to be so, then c
1
(x)u
1
(x) + c
2
(x)u
2
(x) = 0. This is the rst
equation. The second equation is given by plugging u into Lu = f. This
gives p(x)(c
1
u
1
(x) + c
2
(x)u
2
(x)) = f(x). This system of two linear equations
is easily solved. Let w(x) = u
1
(x)u
2
(x) u
2
(x)u
1
(x). The solution is c
1
(x) =
u
2
(x)f(x)/(p(x)w(x)) and c
2
(x) = u
1
(x)f(x)/(p(x)w(x)).
A solution of Lu = f is thus
u(x) =
_
b
x
u
1
(x)u
2
(y)
p(y)w(y)
f(y) dy +
_
x
a
u
2
(x)u
1
(y)
p(y)w(y)
f(y) dy. (8.2)
Furthermore,
u
(x) =
_
b
x
u
1
(x)u
2
(y)
p(y)w(y)
f(y) dy +
_
x
a
u
2
(x)u
1
(y)
p(y)w(y)
f(y) dy. (8.3)
Notice that u(a) = Au
1
(a) and u
(a) = Au
1
(a), while u(b) = Bu
2
(b) and
u
(b) = Bu
2
(b). So this form of the solution is useful for specifying boundary
conditions at a and b.
The general solution is obtained by adding an arbitrary linear combination
C
1
u
1
(x)+C
2
u
2
(x). However often we want a particular solution with boundary
conditions at a and b. Then we use the form above. This can also be written
u(x) = (Kf)(x) =
_
b
a
k(x, y)f(y) dy, (8.4)
where
k(x, y) =
_
u
1
(x)u
2
(y)
p(y)w(y)
if x < y
u
2
(x)u
1
(y)
p(y)w(y)
if x > y
(8.5)
Sometime one thinks of y as a xed source and write the equation
L
x
k(x, y) = (x y). (8.6)
Of course this is just another way of saying that LK = I.
8.4 Second order dierential operators with a
bounded interval: point spectrum
Example. Consider the self-adjoint dierential operator L
D
= d
2
/dx
2
on
L
2
(0, 1) with Dirichlet boundary conditions f(0) = 0 and f(1) = 0 at 0 and
1. Take the solutions u
1
(x) = sin(
x)/
and u
2
(x) = sin(
(1 x))/
.
These are dened in a way that does not depend on which square root of
is taken. (Furthermore, they have obvious values in the limit 0.) Then
p(x) = 1 and w(x) = sin(
)/
. This also does not depend on the cut.

The resolvent is thus ((L
D
)
1
g)(x) = f(x) where
f(x) =
1
sin(
)
__
x
0
sin(
(1 x)) sin(
y)g(y) dy +
_
1
x
sin(
x) sin(
(1 y))g(y) dy
_
.
(8.7)
The spectrum consists of the points = n
2
2
for n = 1, 2, 3, . . .. This is
standard point spectrum. It is amusing to work out the spectral projection at
the eigenvalue n
2
2
. This is the negative of the residue and is explicitly
(P
n
g)(x) = 2
_
1
0
sin(nx) sin(ny)g(y) dy. (8.8)
N
= d
2
/dx
2
on
L
2
(0, 1) with Neumann boundary conditions f
(0) = 0 and f
(1) = 0 at 0 and
1. Take the solutions u
1
(x) = cos(
x) and u
2
(x) = cos(
(1 x)). Then
p(x) = 1 and w(x) =
sin(
). This also does not depend on the cut. The

8.5. SECONDORDER DIFFERENTIAL OPERATORS WITHASEMIBOUNDEDINTERVAL: CONTINUOUS SPECTRUM109
resolvent is thus
((L
N
)
1
g)(x) =
1
sin(
)
__
x
0
cos(
(1 x)) cos(
y)g(y) dy +
_
1
x
cos(
x) cos(
(1 y))g(y) dy
_
.
(8.9)
The spectrum consists of the points = n
2
2
for n = 0, 1, 2, 3, . . .. This is
standard point spectrum. The spectral projection at the eigenvalue n
2
2
for
n = 1, 2, 3, . . . is
(P
n
g)(x) = 2
_
1
0
cos(nx) cos(ny)g(y) dy. (8.10)
For n = 0 it is
(P
0
g)(x) =
_
1
0
g(y) dy. (8.11)
It is interesting to compute the reduced resolvent of L
N
at the eigenvalue 0.
Thus we must compute (L
N
)
1
g, where g = (1 P)g has zero average, and
then let approach zero. This is easy. Expand the cosine functions to second
order. The constant terms may be neglected, since they are orthogonal to g.
This gives
(S
0
g)(x) =
_
x
0
(
1
2
(1 x)
2
+
1
2
y
2
) g(y) dy +
_
1
x
(
1
2
x
2
+
1
2
(1 y)
2
) g(y) dy. (8.12)
From this it is easy to work out that
(S
0
g)(x) =
_
x
0
(
1
2
(1 x)
2
+
1
2
y
2
1
6
)g(y) dy +
_
1
x
(
1
2
x
2
+
1
2
(1 y)
2
1
6
)g(y) dy.
(8.13)
8.5 Second order dierential operators with a
semibounded interval: continuous spectrum
D
= d
2
/dx
2
on
L
2
(0, ) with Dirichlet boundary condition f(0) = 0 at 0. Take the solutions
u
1
(x) = sinh(
x)/
and u
2
(x) = e
x
. Since sinh(iz) = i sin(z), this
is the same u
1
(x) as before. In u
2
(x) the square root is taken to be cut on the
negative axis. Then p(x) = 1 and w(x) = 1. The resolvent is
((L
D
)
1
g)(x) =
1
__
x
0
e
x
sinh(
y)g(y) dy +
_

x
sinh(
x)e
y
g(y) dy
_
.
(8.14)
The spectrum consists of the positive real axis and is continuous.
It is instructive to compute the resolvent of the self-adjoint dierential oper-
ator L
N
= d
2
/dx
2
on L
2
(0, ) with Neumann boundary condition f
(0) = 0
at 0. Again the spectrum consists of the positive real axis and is continuous.
8.6 Second order dierential operators with an
innite interval: continuous spectrum
Example. Consider the self-adjoint dierential operator L = d
2
/dx
2
on L
2
(, ).
There is now no choice of boundary conditions. The resolvent is
((L)
1
g)(x) =
1
2
__
x
x
e
y
g(y) dy +
_

x
e
x
e
y
g(y) dy
_
.
(8.15)
This can also be written in the form
((L )
1
g)(x) =
1
2
|xy|
g(y) dy. (8.16)
The spectrum consists of the positive real axis and is continuous.
8.7 The spectral theorem for normal operators
Throughout the discussion we make the convention that the inner product is
conjugate linear in the rst variable and linear in the second variable.
The great theorem of spectral theory is the following.
Let H be a Hilbert space. Let L be a normal operator. Then there exists
a set K (which may be taken to be a disjoint union of copies of the line) and
a measure on K and a unitary operator U : H L
2
(K, ) and a complex
function on K such that
(ULf)(k) = (k)(Uf)(k). (8.17)
Thus if we write for the operator of multiplication by the function , we get
the representation
L = U
U. (8.18)
The theorem is a generalization of the theorem on diagonalization of normal
matrices. If the measure is discrete, then the norm in the space L
2
() is given
by |g|
2
=

k
[g(k)[
2
(k). The
k
are the eigenvalues of L. The equation
then says
(ULf)
k
=
k
(Uf)
k
. (8.19)
The unitary operator U is given by
(Uf)
k
= (
k
, f), (8.20)
where the
k
are eigenvectors of L normalized so that (k) = 1/(
k
,
k
) .
The inverse of U is given by
U
g =
k
g
k
k
(k). (8.21)
8.7. THE SPECTRAL THEOREM FOR NORMAL OPERATORS 111
The equation
Lf = U
Uf (8.22)
says explicitly that
Lf =
k
(
k
, f)
k
(k). (8.23)
If the measure is continuous, then the norm in the space L
2
(K, ) is given
by |g|
2
=
_
[g(k)[
2
d(k). Then (k)is a function of the continuous parameter
k. The equation then says
(ULf)(k) = (k)(Uf)(k). (8.24)
In quite general contexts the unitary operator U is given by
(Uf)(k) = (
k
, f), (8.25)
but now this equation only makes sense for a dense set of f in the Hilbert space,
and the
k
resemble eigenvectors of L, but do not belong to the Hilbert space,
but instead to some larger space, such as a space of slowly growing functions or
of mildly singular distributions. The inverse of U is given formally by
U
g =
_
g(k)
k
d(k), (8.26)
but this equation must be interpreted in some weak sense. The equation
Lf = U
Uf (8.27)
says formally that
Lf =
_
(k)(
k
, f)
k
d(k). (8.28)
Since the eigenvectors
k
are not in the Hilbert space, it is convenient in
many contexts to forget about them and instead refer to the measure and
the function and to the operators U and U
. The theorem says simply that

every normal operator L is isomorphic to multiplication by a function . The
simplicity and power of the equation L = U
U cannot be overestimated.
The spectral theorem for normal operators says that every normal operator
is isomorphic (by a unitary operator mapping the Hilbert space to an L
2
space)
to a multiplication operator (multiplication by some complex valued function
). The spectrum is the essential range of the function. This is the set of points
such that for each > 0 the set of all points k such that (k) is within of w
has measure > 0. This is obvious; a function (k) has 1/((k) w) bounded if
and only if w is not in the essential range of (k).
8.8 Examples: compact normal operators
The theorem on compact normal operators is a corollary of the general spectral
theorem. Consider a compact normal operator L. Say that it is isomorphic
to multiplication by . Fix > 0 and look at the part of the space where
[L[ . This is just the part of the space that is isomorphic to the part of L
2
where [[ . More explicitly, this is the subspace consisting of all functions g
in L
2
such that g(k) ,= 0 only where [(k)[ . On this part of the space the
operator L maps the unit ball onto a set that contains the ball of radius . Since
L is compact, it follows that this part of the space is nite dimensional. This
shows that the spectrum in the subspace where [L[ is nite dimensional.
Therefore there are only nitely many eigenvectors of nite multiplicity in this
space. Since > 0 is arbitrary, it follows that there are only countably many
isolated eigenvectors of nite multiplicity in the part of the space where [L[ > 0.
In the part of the space where L = 0 we can have an eigenvalue 0 of arbitrary
multiplicity (zero, nite, or innite).
8.9 Examples: translation invariant operators
and the Fourier transform
The nicest examples for the continuous case are given by translation invariant
operators acting in H = L
2
(R, dx). In this case the Fourier transform maps H
into L
2
(R, dk/(2)). The Fourier transform is given formally by
(Ff)(k) = (
k
, f) =
_

e
ikx
f(x) dx. (8.29)
Here
k
(x) = e
ikx
, and we are using the convention that the inner product is
linear in the second variable. The inverse Fourier transform is
(F
1
g)(x) =
_

e
ikx
g(k)
dk
2
. (8.30)
Here are some examples:
Example 1: Translation. Let U
a
be dened by
(T
a
f)(x) = f(x a). (8.31)
Then T
a
is unitary. The spectral representation is given by the Fourier trans-
form. In fact
(FT
a
f)(k) = exp(ika)(Ff)(k). (8.32)
Example 2: Convolution. Let C be dened by
(Cf)(x) =
_

c(x y)f(y) dy =
_

c(a)f(x a) da, (8.33)

8.10. EXAMPLES: SCHR
ODINGER OPERATORS 113

where c is an integrable function. Then C is bounded normal. Then by inte-
grating the rst example we get
(FCf)(k) = c(k)(Ff)(k), (8.34)
where c is the Fourier transform of c.
Example 3: Dierentiation. Let D be dened by
(Df)(x) =
df(x)
dx
. (8.35)
Then D is skew-adjoint. Furthermore, we get
(FDf)(k) = ik(Ff)(k). (8.36)
Notice that the unitary operator in Example 1 may be written
U
a
= exp(aD). (8.37)
Example 4. Second dierentiation. Let D
2
be dened by
(D
2
f)(x) =
d
2
f(x)
dx
2
. (8.38)
Then D is self-adjoint. Furthermore, we get
(FDf)(k) = k
2
(Ff)(k). (8.39)
We can take interesting functions of these operator. For instance (D
2
+
m
2
)
1
is convolution by 1/(2m)e
m|x|
. And exp(tD
2
) is convolution by 1/
2te
x
2
2t
.
8.10 Examples: Schrodinger operators
If V (x) is a real locally integrable function that is bounded below, then
H = D
2
+V (x) (8.40)
is a well-dened self-adjoint operator. Such an operator is called a Schrodinger
operator.
If V (x) as [x[ , then the spectrum of H is point spectrum. Finding
the eigenvalues is a challenge. One case where it is possible to obtain explicit
formulas is when V (x) is a quadratic function.
Also there is an interesting limiting case. If V (x) = 0 for 0 < x < 1 and
V (x) = + elsewhere, then we may think of this as the operator H = D
2
with Dirichlet boundary conditions at the end points of the unit interval. We
know how to nd the spectrum in this case.
If on the other hand, V (x) is integrable on the line, then the spectrum of
H consists of positive continuous spectrum and possibly some strictly negative
eigenvalues. A nice example of this is the square well, where there is a constant
a > 0 with V (x) = a for 0 < x < 1 and V (x) = 0 otherwise. This is another
case where computations are possible.
The calculation of the spectral properties of Schrodinger operators is the
main task of quantum physics. However we shall see that Schrodinger opera-
tors play a role in other contexts as well. (One example will be in calculus of
variations.)
8.11 Subnormal operators
The spectral theorem for normal operators is a landmark. However not every
operator is normal. One important class of operators with fascinating spectral
properties consists of the subnormal operators. A subnormal operator is an
operator that is the restriction of a normal operator to an invariant subspace.
First consider the case of bounded operators. An operator S : H H is
subnormal if there exists a larger Hilbert space H
with H H
as a closed
subspace, a normal operator N : H
that leaves H invariant, and such

that N restricted to H is S.
Example: Let H = L
2
([0, ), dt). For each a 0 dene
(S
a
f)(t) = f(t a) (8.41)
for a t and
(S
a
f)(t) = 0 (8.42)
for 0 t < a. Then S
a
is subnormal.
To see this, consider the bigger space H
= L
2
(R, dt) and consider H as
the subspace of functions that vanish except on the positive reals. Let U
a
be
translation by a on H
. If a 0, then the subspace H is left invariant by U

a
.
Then S
a
is U
a
acting in this subspace.
It follows from the spectral theorem that a subnormal operator S is isomor-
phic by a unitary operator U : H M to a multiplication operator that sends
g(k) to (k)g(k). Here M is a closed subspace of an L
2
space. For each f in H
we have Uf in M and USf(k) = (k)(Uf)(k) in M.
Now it is more dicult to characterize the spectrum. A number w is in the
resolvent set if 1/((k) w)g(k) is in M for every function g(k) in M. However
it is not sucient that this is a bounded function. If, for instance, every function
g(k) in M has an extension to an analytic function g(z) dened on some larger
region, then one would want 1/((z) w)g(z) to also be an analytic function
in this region. So we need to require also that w is not in the range of the
extension (z).
Example: Let F be the Fourier transform applied to the Hilbert space H
of L
2
functions that vanish except on the positive axis. Then the image of this
transform consists of the subspace M of L
2
functions g(k) that are boundary
values of analytic functions g(z) in the lower half plane. The operator S
a
for a >
0 is isomorphic to multiplication by (k) = exp(iak) acting in this subspace.
This function extends to a function (z) = exp(iaz) dened in the lower half
8.11. SUBNORMAL OPERATORS 115
plane. So the spectrum is the range of this function. But the image of z 0
under exp(iaz) is the unit circle [w[ 1. So this is the spectrum.
The adjoint of a subnormal operator S : H H is another operator S
:
H H. The adjoint S
need not be subnormal.

Lemma. If N : H
is normal, H H
, and S : H H is the
restriction of N to H, then S
: H H is given by S
= PN
, where P is the
orthogonal projection of H
onto H.
Proof: If u is in H, for each v in H we have
(S
u, v) = (u, Sv) = (u, Nv) = (N
u, v). (8.43)
This says that N
uS
u is orthogonal to every v in H. Since S
u is in H, this
implies that S
u is the orthogonal projection of N
u onto H.
Theorem. If S is a subnormal operator, then SS
S as quadratic forms,
that is, (u, SS
u) (u, S
Su) for all u in H.

Proof: First note that for u in H
we have
(N
u, N
u) = (u, NN
u) = (u, N
Nu) = (Nu, Nu). (8.44)

Then for u in H we have
(S
u, S
u) = (PN
u, PN
u) (N
u, N
u) = (Nu, Nu) = (Su, Su). (8.45)

Corollary. If S is a subnormal operator and Su = 0, then S
u = 0. Thus
the null space of S is contained in the null space of S
.
Corollary. If S is a subnormal operator and Su = u, then S
u =

u. Thus
every eigenvector of S is an eigenvector of S
.
It is not true that every eigenvalue of S
is an eigenvalue of S. It is more
typical that S
has eigenvalues while S does not. In fact we shall see examples

in which S
has anomalous point spectrum, while S has residual spectrum.

Theorem. If the Hilbert space is nite dimensional, then every subnormal
operator is normal.
Proof: Let S be a subnormal operator acting in H. Since the space H is
nite dimensional, S has an eigenvector u in H with Su = u. Since S is
subnormal, it follows that S
u =

u. Let v be a vector in H that is orthogonal
to u. Then (Sv, u) = (v, S
u) =

(v, u) = 0. Thus the orthogonal complement
of u in H is also an invariant space, so the operator S restricted to this smaller
space is also subnormal. Continue in this way until one nds an orthogonal
basis of eigenvectors for S.
There are examples in which one might want to consider unbounded sub-
normal operators. One possible denition might be the following. Consider an
unbounded normal operator N. Thus there is a dense domain D(N) H
such
that N : D(N) H
is normal. Let H be a Hilbert space that is a closed sub-

space of H
. Suppose that D(N) H is dense in H and that N sends vectors in

this dense subspace into H. Then if D(S) = D(N) H and S is the restriction
of N to D(S), the operator S is subnormal. Then D(S) is dense in H, and
S : D(S) H is a closed, densely dened operator.
Example: Let H = L
2
([0, ), dt). Dene
(Sf)(t) = f
(t) (8.46)
on the domain consisting of all f in H such that f
is also in H and f(0) = 0.

Then S is subnormal.
To see this, consider the bigger space H
= L
2
(R, dt) and consider H as
the subspace of functions that vanish except on the positive reals. Let N be
dierentiation on H
. Notice that if f is in D(N) and is also in H, then f(0) = 0

automatically.
If N is a normal operator acting in H
, so N : D(N) H
, then its adjoint

N
is also a normal operator, and in fact D(N
) = D(N). So if S : D(S) H
is subnormal, then D(S) = D(N) H = D(N
) H. Furthermore, if u is in
D(N
) H, then u is in D(S
) and S
u = PN
u. This can be seen from the

computation (u, Sv) = (u, Nv) = (N
u, v) for all v in D(S).

We conclude that for a subnormal operator D(S) D(S
) and (S
, S
u)
(Su, Su) for all u in D(S). This is also an easy computation: (S
u, S
u) =
(PN
u, PN
u) (N
u, N
u) = (Nu, Nu) = (Su, Su).

Example: For the operator S in the last example we have
(S
f)(t) = f
(t) (8.47)
on the domain consisting of all f in H such that f
is also in H. There is no
boundary condition at zero. This is a case where D(S) D(S
) and the two

domains are not equal.
8.12 Examples: forward translation invariant op-
erators and the Laplace transform
The example in the last section is the operator theory context for the theory of
the Laplace transform.
The Laplace transform of a function f in L
2
([0, ), dt) is
(Lf)(z) =
_

0
e
zt
f(t) dt. (8.48)
If we think of z = i on the imaginary axis, then when regarded as a function
of this is the Fourier transform . However it extends as an analytic function
to in the lower half plane, that is, to z in the right half plane.
Let a 0 and let S
a
be the operator of right translation lling in with zero
as dened in the previous section. Then
(LS
a
f)(z) =
_

a
e
zt
f(t a) dt = e
az
(Lf)(z). (8.49)
So S
a
is isomorphic to multiplication by e
az
acting on the Hilbert space of
functions analytic in the right half plane. Its spectrum consists of the closed
unit disk.
8.12. EXAMPLES: FORWARDTRANSLATIONINVARIANT OPERATORS ANDTHE LAPLACE TRANSFORM117
Let c be an integrable function dened on the positive axis, and let the causal
convolution C be dened by
(Cf)(t) =
_
t
0
c(t u)f(u) du =
_
t
0
c(a)f(t a) da. (8.50)
Then C is subnormal, and
(LCf)(z) = c(z)(Lf)(z), (8.51)
where c is the Laplace transform of c.
Let D
0
= d/dt be dierentiation with zero boundary conditions at the origin.
Then
(LD
0
f)(z) = z(Lf)(z). (8.52)
Notice that the boundary condition is essential for integration by parts. The
spectrum of D
0
consists of the closed right half plane. We can write exp(aD
0
) =
S
a
for a 0. This operator satises the dierential equation dS
a
f/da =
D
0
S
a
f for a 0, provided that f is in the domain of D
0
(and in particu-
lar satises the boundary condition).
The adjoint of a subnormal operator need not be subnormal. Thus, for
instance, the adjoint of S
a
is
(S
a
f)(t) = f(t +a). (8.53)
Its spectrum is also the closed unit disc, but the interior of the disk consists of
point spectrum. Notice that the Laplace transform does not send this into a
multiplication operator. In fact,
(LS
a
f)(z) =
_

0
e
zt
f(t +a) dt = e
az
_
(Lf)(z)
_
a
0
e
zt
f(t) dt
_
. (8.54)
Similarly, the adjoint of D
0
is D, where D = d/dt with no boundary
condition. Again the Laplace transform does not make this into a multiplication
operator. In fact, we have
(LDf)(z) = z(Lf)(z) f(0). (8.55)
The spectrum of D consists of the closed left half plane. The interior consists
of point spectrum. We can write exp(aD) = exp(aD
0
) = S
a
for a 0. Even
though D does not have a spectral representation as a multiplication operator,
this operator satises the dierential equation dS
a
f/da = DS
a
f for a 0,
provided f is in the domain of D.
Example: Solve the dierential equation (D + k)f = g, with k > 0, with
boundary condition f(0) = c. The operator D +k is not invertible, since k is
an eigenvalue of D. However we can write f = u ce
kt
and solve
(D
0
+k)u = g. (8.56)
The solution is u = (D
0
+k)
1
g. In terms of Laplace transforms this is u(z) =
1/(z +k) g(z). It follows that u is given by the causal convolution
u(t) =
_
t
0
e
k(tu)
g(u) du. (8.57)
Thus
f(t) = f(0)e
kt
+
_
t
0
e
k(tu)
g(u) du. (8.58)
8.13 Quantum mechanics
Here is a dictionary of the basic concepts. Fix a Hilbert space. A quantum
observable is a self-adjoint operator L. A quantum state is a unit vector u. The
expectation of the observable L in the state u in D(L) is
= (u, Lu). (8.59)
The variance of the observable L in the state u in D(L) is
2
= |(L I)u|
2
. (8.60)
Here are some observables. The Hilbert space is L
2
(R, dx). The momentum
observable is p = ihd/dx. Here h > 0 is Plancks constant. The position
observable is q which is multiplication by the coordinate x. The Heisenberg
uncertainty principle says that for every state the product
p
q
h/2.
If L is an observable and f is a real function, then f(L) is an observable.
Take the case where the function is 1
A
, the indicator function of a set A of
position coordinate values. Then 1
A
(q) has expectation
(u, 1
A
(q)u) =
_
A
[u(x)[
2
dx. (8.61)
This is the probability that the position is in A. Similarly, take the case where
the function is 1
B
, the indicator function of a set B of momentum coordinate
values. Since in the Fourier transform representation p is represented by multi-
plication by hk, it follows that 1
B
(p) has expectation
(u, 1
B
(p)u) =
_
{k| hkB}
[ u(k)[
2
dk
2
. (8.62)
This is the probability that the momentum is in B.
Energy observables are particularly important. The kinetic energy observable
is H
0
= p
2
/(2m) = h
2
/(2m)d
2
/dx
2
. Here m > 0 is the mass. The spectrum of
H
0
is the positive real axis. The potential energy observable is V = v(q) which is
multiplication by v(x). Here v is a given real function that represents potential
energy as a function of the space coordinate. The spectrum of V is the range
of the function v. The total energy observable or quantum Hamiltonian is
H = H
0
+V =
p
2
2m
+v(q) =
h
2
2m
d
2
dx
2
+v(x). (8.63)
8.14. PROBLEMS 119
If we assume that the function v is bounded, then H is a self-adjoint operator.
(In many cases when v is only bounded below it remains a self-adjoint oper-
ator.) The problem of investigating its spectrum is of the utmost importance
for quantum mechanics. Since H
0
and V do not commute, this is not an easy
problem.
Suppose that the total energy observable H is a self-adjoint operator. Then
the time evolution operator is the unitary operator exp(itH/h). A central
problem of quantum mechanics is to compute this operator. This is not easy,
because while exp(itH
0
/h) and exp(itV/h) are easy to compute, the opera-
tors H
0
and V do not commute. So there is no direct algebraic way to express
exp(itH/h) in terms of the simpler operators. Nevertheless, we shall encounter
a beautiful formula for this time evolution operator in the next chapter.
8.14 Problems
1. A particularly fascinating self-adjoint operator is the quantum harmonic
oscillator. (This operator also occurs in disguised form in other contexts.)
It is
N =
1
2
_
d
2
dx
2
+x
2
1
_
acting in L
2
(, ). Show that it factors as
N = A
A,
where
A =
1
2
_
x +
d
dx
_
and
A
=
1
2
_
x
d
dx
_
.
2. Show that AA
= A
A+I.
3. Solve the equation Au
0
= 0. Show that Nu
0
= 0.
4. Show that if Nu
n
= nu
n
and u
n+1
= A
u
n
, then Nu
n+1
= (n + 1)u
n+1
.
Thus the eigenvalues of N are the natural numbers. These are the stan-
dard type of point spectrum.
5. Show that each eigenfunction u
n
is a polynomial in x times u
0
(x). Find
the polynomials for the cases of n = 0, 1, 2, 3 explicitly (up to constant
factors). Verify that each u
n
belongs to the Hilbert space.
6. It may be shown that A
is a subnormal operator and so A is the adjoint

of a subnormal operator. Find all eigenvalues (point spectrum) of A.
Find each corresponding eigenvector. Verify that it belongs to the Hilbert
space.
7. Find all eigenvalues (point spectrum) of A
. Find the spectrum of A
.
What kind of spectrum is it?
8. If A
is indeed a subnormal operator, then we should have A
A AA
as quadratic forms. Is this the case?

Chapter 9
Calculus of Variations
9.1 The Euler-Lagrange equation
The problem is to nd the critical points of
F(y) =
_
x
2
x
1
f(y, y
x
, x) dx. (9.1)
The dierential of F is
dF(y)h =
_
x
2
x
1
(f
y
h +f
y
x
h
x
) dx =
_
x
2
x
1
(f
y

d
dx
f
y
x
)hdx +f
y
x
h [
x
2
x
1
. (9.2)
Thus for the dierential to be zero we must have the Euler-Lagrange equation
f
y

d
dx
f
y
x
= 0. (9.3)
This is an equation for the critical function y. It is second order, and it has the
explicit form
f
y
f
yy
x
dy
dx
f
y
x
y
x
d
2
y
dx
2
f
xy
x
= 0. (9.4)
This equation is linear in d
2
y/dx
2
. However the coecients are in general
nonlinear expressions in y, dy/dx, and x.
If the y are required to have xed values y = y
1
at x = x
1
and y = y
2
at
x = x
2
at the end points, then the h are required to be zero at the end points.
The corresponding boundary term is automatically zero.
If the y and h are free to vary at an end point, then at a critical point one
must have f
y
x
= 0 at at the end points.
Sometimes one wants to think of
F
y(x)
=
f
y

d
dx
f
y
x
(9.5)
121
122 CHAPTER 9. CALCULUS OF VARIATIONS
as the gradient of F, where the inner product is given by the integral. This
expression is then known as the variational derivative. The Euler-Lagrange
equation then says that the variational derivative is zero.
Example. Consider the problem of minimizing the length
F(y) =
_
x
2
x
1
_
1 +y
2
x
dx (9.6)
between the points (x
1
, y
1
) and (x
2
, y
2
). The boundary condition are y(x
1
) = y
1
and y(x
2
) = y
2
. The solution of the Euler-Lagrange equation is y
x
= C, a
curve of constant slope. So the solution is y y
1
= m(x x
1
), where m =
(y
2
y
1
)/(x
2
x
1
).
Example. Consider the problem of minimizing the length
F(y) =
_
x
2
x
1
_
1 +y
2
x
dx (9.7)
between the lines x = x
1
and x = x
2
. The boundary conditions for the Euler-
Lagrange equation are y
x
(x
1
) = 0 and y
x
(x
2
) = 0. The solution of the Euler-
Lagrange equation is y
x
= C, a curve of constant slope. But to satisfy the
boundary conditions C = 0. So the solution is y = A, where A is a arbitrary
constant.
9.2 A conservation law
Say that one has a solution of the Euler-Lagrange equation. Then
d
dx
f = f
y
y
x
+f
y
x
y
xx
+f
x
=
d
dx
(y
x
f
y
x
) +f
x
. (9.8)
Thus
d
dx
H +f
x
= 0, (9.9)
where
H = y
x
f
y
x
f. (9.10)
If f
x
= 0, then this says that H is constant. This is the conservation law. It is
a nonlinear equation for y and dy/dx. It takes the explicit form
y
x
f
y
x
f = C. (9.11)
Thus it is rst order, but in general fully nonlinear in y
x
. How in practice does
one solve a problem in calculus of variations? There is a way when the function
f(y, y
x
) does not depend on x. Use the conservation law H(y, dy/dx) = C
to solve for dy/dx = (y). Perhaps this equation can be solved directly. Or
write dx = dy/(y). Integrate both sides to get x in terms of y. Or make a
substitution expressing y in terms of a new variable u, and get x and y each in
terms of u.
9.3. SECOND VARIATION 123
Example 1: This example comes from minimizing the area of a surface of
revolution. The unknown is a function y of x between a and a. The value of
y at a is r. The element of area is proportional to 2y ds = 2y
_
1 +y
2
x
dx.
Let f(y, y
x
) = 2y
_
1 +y
2
x
. Then H = 2y/
_
1 +y
2
x
= C. The dierential
equation to be solved is (dy/dx)
2
= (k
2
y
2
1), where k = 2/C. This has
solution y = (1/k) cosh(k(x x
0
)). By symmetry x
0
= 0. So we need to solve
rk = cosh(ak). This equation has to be solved for k. Fix r and vary a. When a
is small enough, then cosh(ak) cuts the line rk in two points. Thus there are two
solutions of the Euler-Lagrange equations, corresponding to y = (1/k) cosh(kx)
with the two values of k. The derivative is dy/dx = sinh(kx). The smaller value
of k gives the smaller derivative, so this is the minimum area surface satisfying
the boundary conditions.
Example 2: This example comes up nding the maximum area for given
arc length. For simplicity consider y as a function of x between a and a,
with value 0 at the two end points. For simplicity let the arc length satisfy
2a < < a. The area is the integral from a to a of y dx, while the length
is the integral from a to a of ds =
_
1 +y
2
x
dx. This is a Lagrange multiplier
problem. The function is f(y, y
x
) = y
_
1 +y
2
x
. The conserved quantity
is H = 1/
_
1 +y
2
x
y = c
1
. The dierential equation is dx/dy = (y
c
1
)/
_
(
2
(y c
1
)
2
. Thus x c
2
=
_
2
(y c
1
)
2
, which is the equation
of a circle (x c
2
)
2
+ (y c
1
)
2
=
2
. The Lagrange multiplier turns out to be
the radius of the circle. By symmetry c
2
= 0. Furthermore c
1
=
2
a
2
. If
we let sin() = a/, then the equation sin() < < (/2) sin() translates into
2a < = 2 < a.
9.3 Second variation
Say that y is a solution of the Euler-Lagrange equation that has xed values
at the end points. Then the second dierential of F is obtained by expanding
F(y +h) = F(y) +dF(y)h +
1
2
d
2
F(y)(h, h) + . The result is
d
2
F(y)(h, h) =
_
x
2
x
1
(f
yy
h
2
+ 2f
yy
x
hh
x
+f
y
x
y
x
h
2
x
) dx. (9.12)
The functions h have value 0 at the end points, so we may freely integrate by
parts. The result may thus also be written
d
2
F(y)(h, h) =
_
x
2
x
1
((f
yy

d
dx
f
yy
x
)h
2
+f
y
x
y
x
h
2
x
) dx. (9.13)
Yet another form is
d
2
F(y)(h, h) =
_
x
2
x
1
[
d
dx
(f
y
x
y
x
h
x
) + (f
yy

d
dx
f
yy
x
)h]hdx = (Lh, h). (9.14)
We recognize
Lh =
d
dx
(f
y
x
y
x
h
x
) + (f
yy

d
dx
f
yy
x
)h (9.15)
as a Sturm-Liouville operator. In fact, if f(y, y
x
, x) =
1
2
y
2
x
+g(y, x), then
Lh = h
xx
+g
yy
h (9.16)
is a Schrodinger operator. The coecient g
yy
has the solution of the Euler-
Lagrange inserted, so it is regarded as a function of x. The operator has Dirichlet
boundary conditions at the end points of the interval from x
1
to x
2
.
9.4 Interlude: The Legendre transform
The Legendre transform of a function L of v is dened as follows. Let p = dL/dv.
Find the inverse function dening v as a function of p. Then the Legendre
transform is a function H of p satisfying dH/dp = v. The constant of integration
is chosen so that H(0) +L(0) = 0.
Example: Let L = e
v
v 1. Then the derivative is p = e
v
1. The inverse
function is v = log(1 +p). The integral is H = (1 +p) log(1 +p) p.
There is a remarkable formula relating the Legendre transform of a function
to the original function. Let L be a function of v. Dene
p =
dL
dv
. (9.17)
Then
H = pv L. (9.18)
is the Legendre transform of L. In the following we want to think of L as a
function of the v, while H is a function of the dual variable p. The variable p
is covariant and is dual to the variable v, which is contravariant. Thus in this
formula v is dened in terms of p by solving the equation p = dL/dv for v.
The fundamental theorem about the Legendre transform is
v =
dH
dp
. (9.19)
Proof: Let H = pv L as above By the product rule and the chain rule and
the denition of p we get
dH
dp
= v +p
dv
dp

dL
dv
dv
dp
= v +p
dv
dp
p
dv
dp
= v. (9.20)
Thus the situation is symmetric, and one can go back from H to L in the
same way as one got from L to H. Conclusion: Two functions are Legendre
transforms of each other when their derivatives are inverse functions to each
other.
One would like a condition that guarantees that the equations that dene
the Legendre transform actually have solutions. In order for p = dL/dv to have
a unique solution v, it would be useful to have dL/dv to be strictly increasing.
This is the same as saying that dp/dv = d
2
L/dv
2
> 0. The inverse function
then has derivative given by the inverse dv/dp = d
2
H/dp
2
> 0.
9.4. INTERLUDE: THE LEGENDRE TRANSFORM 125
Example: Let L = e
v
v1. Then we have seen that H = (p+1) log(p+1)p.
This also follows from the general formula H = pv L. In fact, since p = e
v
1
has inverse v = log(1+p), we have H = pv L = p log(1+p) [(1+p) log(1+
p) 1] = (1 + p) log(1 + p) p. In this example the second derivatives are e
v
and 1/(p + 1). These are both positive, and they are reciprocals of each other.
Example: Let L = v
a
/a with 1 < a < dened for v 0. Then H = p
b
/b
with 1 < b < dened for p 0. In this case p = v
a1
and v = p
b1
. The
relation between a and b is (a 1)(b 1) = 1. This may also be written as
1/a + 1/b = 1. Thus a and b are conjugate exponents. The second derivatives
are (a 1)v
a2
and (b 1)p
b2
. Again they are reciprocals.
The Legendre transform plays a fundamental role in thermodynamics and
in mechanics. Here are examples from mechanics involving kinetic energy.
Example: Start with L = (1/2)mv
2
. Then p = mv. So v = p/m and
H = pv L = p
2
/(2m). This H is the Legendre transform of L.
Example: If we start instead with H = p
2
/(2m), then v = p/m. So p = mv
and L = pv H = (1/2)mv
2
. The Legendre transform of L brings us back to
the original H.
Example: A famous relativistic expression for energy is H =
_
m
2
c
4
+p
2
c
2
.
It is an amusing exercise to compute that L = mc
2
_
1 v
2
/c
2
. The key is
the relation p = mv/
_
1 v
2
/c
2
between the derivatives.
The Legendre transform has a generalization to several dimensions. Let L
be a function of v
1
, . . . , v
n
. Dene
p
j
=
L
v
j
, (9.21)
so dL =

j
p
j
dv
j
. (This is the dierential of L, so it is a one-form.) Then
H =
k
p
k
v
k
L. (9.22)
is the Legendre transform of L. In the following we want to think of L as a
function of the v
j
, while H is a function of the dual variables p
k
. The variables p
k
are covariant and dual to the variable v
j
, which are contravariant. Thus in this
formula the v
j
are dened in terms of p
k
by solving the equation p
k
= L/v
k
for the v
j
.
The fundamental theorem about the Legendre transform in this context is
v
k
=
H
p
k
. (9.23)
Proof: Let H =

k
p
k
v
k
L as above By the product rule and the chain
rule and the denition of p we get
H
p
j
= v
j
+
k
p
k
v
k
p
j
k
L
v
k
v
k
p
j
= v
j
+
k
p
k
v
k
p
j
k
p
k
v
k
p
j
= v
j
. (9.24)
Thus the situation is symmetric, and one can go back from H to L in the
same way as one got from L to H. Conclusion: Two functions are Legendre
transforms of each other when their derivatives are inverse functions to each
other.
One would like a condition that guarantees that the equations that dene
the Legendre transform actually have solutions. In order for the equation to
have a solution locally, the matrix
p
k
v
j
=

2
L
v
k
v
j
(9.25)
should be invertible. A particularly nice condition is that this matrix is strictly
positive denite. The inverse matrix is then
v
j
p
k
=

2
H
p
j
p
k
. (9.26)
It is then also strictly positive denite.
9.5 Lagrangian mechanics
One is given a Lagrangian function L that is a function of position q and velocity
q and possibly time t. The problem is to nd the critical points of
S(q) =
_
t
2
t
1
L(q, q
t
, t) dt. (9.27)
The function q represents position as a function of time. It has xed values q
1
and q
2
at the end points t
1
and t
2
. The dierential of S is
dS(q)h =
_
t
2
t
1
(L
q
h +L
q
h
t
) dt =
_
t
2
t
1
(L
q

d
dt
L
q
)hdt. (9.28)
Here h is a function with values 0 at the end points. Thus for the dierential
to be zero we must have the Euler-Lagrange equation
L
q

d
dt
L
q
= 0. (9.29)
Say that one has a solution of the Euler-Lagrange equation. Then
d
dt
L = L
q
q
t
+L
q
q
tt
+L
t
=
d
dt
(q
t
L
q
) +L
t
. (9.30)
Thus along a solution
d
dt
H +
L
t
= 0, (9.31)
where
H = q
t
L
q
L. (9.32)
If L
t
= 0, then this says that H is constant. This is the energy conservation
law.
9.6. HAMILTONIAN MECHANICS 127
Dene
p =
L
q
. (9.33)
This is the momentum variable. Then the Euler-Lagrange equation says that
along a solution
dp
dt
=
L
q
. (9.34)
9.6 Hamiltonian mechanics
Let
H = p q L. (9.35)
be the Legendre transform of L, where q is dened in terms of p implicitly by
p = L/ q.
In the following we want to think of L as a function of q and q and possibly
t, while H is a function of q and p and possibly t. The momentum variable p is
covariant and is dual to the velocity variable q, which is contravariant. However
p is expressed in terms of q and q by solving the Legendre transform equation
p = L/ q. According to the properties of the Legendre transform we have
q =
H
p
. (9.36)
In fact, the proof is simple if we remember that q is xed:
H
p
= q +p
q
p

L
q
q
p
= q. (9.37)
It also follows that
H
q
=
L
q
. (9.38)
This also has an easy proof if we remember that p is xed and so q depends on
q:
H
q
= p
q
q

L
q

L
q
q
q
=
L
q
. (9.39)
The Euler-Lagrange equation says that along a solution we have
dp
dt
=
H
q
. (9.40)
Since p = L/ q determines the function p in terms of q, dq/dt, and t, it follows
that along a solution
dq
dt
=
H
p
(9.41)
determines the function dq/dt in terms of q, p, and t. The last two equations
are Hamiltons equations. They are a rst order system, linear in dq/dt and
dp/dt, but nonlinear in q, p, t.
It is easy to check from Hamiltons equations that along a solution
dH
dt
=
H
t
. (9.42)
When H does not depend explicitly on t, then this has an integral in which H
is constant. This integral is conservation of energy. It gives a situation in which
p is related to q. This allows an equation in which dq/dt is related to q. This
relation may be highly nonlinear.
9.7 Kinetic and potential energy
The most classical situation is when
L =
1
2
m q
2
V (q). (9.43)
The Euler-Lagrange equation is just.
V
(q)
dq
t
dt
= 0. (9.44)
This is Newtons law of motion.
In this problem the momentum is p = m q. The Hamiltonian is
H =
1
2m
p
2
+V (q). (9.45)
Let us look at this from the point of view of maximum and minimum. The
second dierential is determined by the Schrodinger operator
L = m
d
2
dt
2
V
(q(t)). (9.46)
Here q(t) is a function of t satisfying the Euler-Lagrange equation and the
boundary conditions. So for instance if this operator with Dirichlet boundary
conditions at t
1
and t
2
has strictly positive eigenvalues, then the solution will
be a minimum. This will happen, for instance, if V
(q) 0. In many other

cases, the solution will not be a minimum of the action, but only a stationary
point.
Example: Take an example with V
(q) < 0. Then the problem is to nd q(t)

that minimizes the integral from t
1
to t
1
of (1/2)m q
2
V (q). Here q(t
1
) = q
1
and q(t
2
) = q
2
are xed. Clearly there is tradeo. One would like the solution
to linger as long as possible in the region where V (q) is large. On the other
hand, to satisfy the boundary conditions the solution should not move too fast.
The solution will start at the point y
1
at time t
1
. Then it go reasonably, but not
excessively, rapidly to the region where V (q) is large. There it will linger. Then
it will fall back to the point y
2
, arriving there at time t
2
. This minimization
problem seems to have nothing to do with Newtons laws. But it gives the same
answer.
9.8. PROBLEMS 129
9.8 Problems
Let L be a function of q and q given by
L =
1
2
m q
2
V (q).
Let
S(q) =
_
t
2
t
1
Ldt,
with functions q of t satisfying q = q
1
at t = t
1
and q = q
2
at t = t
2
.
1. Show that
dS(q)h = (m
d
dt
q V
(q), h),
where h satises Dirichlet boundary conditions at t
1
and t
2
.
2. Consider a q(t) for which dS(q) = 0. Show that
d
2
S(q)(h, h) =
_
t
2
t
1
[m
_
dh
dt
_
2
V
(q)h
2
] dt.
where the functions h satisfy Dirichlet boundary conditions at t
1
and t
2
.
3. Consider a q(t) for which dS(q) = 0. Show that
d
2
S(q)(h, h) = (h, [m
d
2
dt
2
V
(q)]h),
where the operator satises Dirichlet boundary conditions at t
1
and t
2
.
4. Show that if V (q) is concave down, then the solution q(t) of the variational
problem is actually a minimum.
5. Let H = m q
2
L = (1/2)m q
2
+V (q). Show that H = E along a solution,
where E is a constant.
6. From now on take the example V (q) = (1/2)kq
2
. Here k > 0. Note the
sign. We are interested in solutions with E > 0. Let =
_
k/m. Show
that q = C sinh(t) is a solution, and nd the constant C in terms of E.
7. Take t
1
= T and t
2
= T. Take q
1
= a and q
2
= a. Fix a. Write the
boundary condition a = C sinh(T) as a relation between T and E. Show
that T 0 implies E , while T implies E 0.
8. Interpret the result of the last problem intuitively in terms of particle
motion satisfying conservation of energy.
9. Interpret the result of the same problem intuitively in terms of a mini-
mization problem.
9.9 The path integral
Let us return to the problem of evaluating the quantum mechanical time evo-
lution exp(iitH/h). Here the total energy operator H = H
0
+ V is the sum
of two non-commuting operators, corresponding to potential energy and kinetic
energy.
The problem is easy when we just have kinetic energy. Then we must
evaluate exp(itH
0
/h), where H
0
= p
2
/(2m) = h
2
/(2m)d
2
/dx
2
. In the
Fourier transform representation this is multiplication by h
2
/(2m)k
2
. So in the
Fourier transform representation the unitary time evolution is multiplication by
exp(it( h/m)k
2
/2). This is like the heat equation with complex diusion coef-
cient i
2
= i(h/m). So the solution is convolution by the inverse Fourier trans-
form, which is 1/
2i
2
t exp(x
2
/(2i
2
t)) = 1/
_
2i(h/m)t exp(ix
2
/(2(h/m)t).
In other words,
(exp(
itH
0
h
)u)(x) =
_

exp(
im(x x
)
2
th
)u(x
)
dx
_
2i( h/m)t
. (9.47)
The calculation of the exponential is also easy when we just have potential
energy. In fact,
(exp(
itV
h
)u)(x) = exp(
itv(x)
h
)u(x). (9.48)
The Trotter product formula is a remarkable formula that expresses the
result for the sum H = H
0
+ V in terms of the separate results for H
0
and for
V . It may be proved under various circumstances, for example when V is a
bounded operator. The formula says that
exp(
itH
h
)u = lim
n
_
exp(
i(t/n)H
0
h
) exp(
i(t/n)V
h
)
_
n
u. (9.49)
We can write out the Trotter product formula in detail using the results
obtained before for H
0
and for V separately. Write t = t/n. Then the quantity
of interest is (exp(
itH
h
)u)(x) = u(x, t) given by the Trotter formula. This works
out to be
u(x, t) = lim
n
_

exp
_
_
_
_
i
n1
j=0
_
_
x
j+1
x
j
t
_
2
V (x
j
)
_
t
h
_
_
_
_
u(x
0
)
dx
n1
dx
0
(
_
2i( h/m)t)
n
,
(9.50)
where x
n
= x.
So far this has been rigorous. However now we follow Feynman and take the
formal limit as n and t 0 with n = t xed. This gives
(exp(
itH
h
)u)(x) =
_
exp
_
_
_
_
i
_
t
0
_
_
dx(t
)
dt
_
2
V (x(t
))
_
dt
h
_
_
_
_
u(x
0
) Tx,
(9.51)
9.10. APPENDIX: LAGRANGE MULTIPLIERS 131
where the integral is over all paths x from some x
0
at time 0 to xed nal x at
time t. This can also be written
(exp(
itH
h
)u)(x) =
_
exp
_
i
_
t
0
L(x(t
),
dx(t
)
dt
) dt
h
_
u(x
0
) Tx, (9.52)
where
L(x(t
),
dx(t
)
dt
) =
_
dx(t
)
dt
_
2
V (x(t
)) (9.53)
is the Lagrangian. Thus the integrand in the path integral is the exponential of
i times the action divided by h.
This expression goes some way toward an explanation of the principle of
stationary action. Consider a part of the integral near a path that is not a
critical point of the action. Then nearby paths will have considerably dierent
values of the action, and after division by h the phase will be very dierent. So
there will be a lot of cancelation. On the other hand, the part of the integral
near a path that is a critical point will contribute more or less the same value
of the action. So there will be no cancelation.
9.10 Appendix: Lagrange multipliers
In this section y will represent the coordinates y
1
, . . . , y
n
. We are interested in
nding the critical points of a scalar function F(y) = F(y
1
, . . . , y
n
).
Consider rst the problem of nding a critical point of F(y) with no con-
straints. The usual method is to compute the dierential and set it equal to
zero, so that the equation to be solved is dF(y) = 0.
Consider the problem of nding a critical point of F(y) subject to a con-
straint G(y) = .
The method of Lagrange multipliers is to say that the dierential of F(y)
without the constraint must satisfy
dF(y) = dG(y) (9.54)
for some . That is, the function can vary only by relaxing the constraint. The
change in the function must be proportional to the change in the constraint.
This fundamental equation is an equation for dierential forms, and it takes the
same form in every coordinate system.
Example: Maximize y
1
y
2
subject to y
2
1
+ y
2
2
= . The equation is y
2
dy
1
+
y
1
dy
2
= (2y
1
dy
1
+2y
2
dy
2
). This gives y
2
= 2y
1
and y
1
= 2y
2
. Eliminating
we get y
2
1
= y
2
2
. Combine this with y
2
1
+ y
2
2
= . We get y
1
=
_
/2 and
y
2
=
_
/2. The value of y
1
y
2
at these points are /2. If you care, you can
also compute that = 1/2.
Say that y is a critical point, so in particular it satises the constraint G( y) =
. Then dG( y) = d. It follows that dF( y) = dG( y) = d. The conclusion is
that the Lagrange multiplier is the rate of change of the value of the function
at the critical point as a function of the constraint parameter, that is, =
dF( y)
d
.
Example: In the last example the derivative of the maximum value /2
with respect to is = 1/2.
In calculations, one wants to maximize F(y) subject to the constraint G(y) =
. The idea is to take arbitrary. Set

F(y) = F(y) G(y) and require
d
F(y) = 0. This gives the above equation for y and . However one also has
the constraint equation for y and . These two are then solved simultaneously.
The method of Lagrange multipliers extends to the case when there are
several constraints.
Consider the problem of nding a critical point of F(y) subject to constraint
G
1
(y) =
1
, . . . , G
m
(y) =
m
..
The method of Lagrange multipliers is to say that the dierential of F(y)
without the constraint must satisfy
dF(y) =
m
j=1
j
dG
j
(y) (9.55)
for some . That is, the function can vary only by relaxing the constraint. The
change in the function must be proportional to the change in the constraints.
This fundamental equation is an equation for dierential forms, and it takes the
same form in every coordinate system.
Say that y is a critical point, so in particular it satises the constraints
G
j
( y) =
j
. Then dG
j
( y) = d
j
. It follows that dF( y) =

j

j
dG
j
( y) =
j

j
d
j
. The conclusion is that the Lagrange multiplier
j
is the rate of change
of the value of the function at the critical point as a function of the constraint
parameter
j
, with the other constraint parameters xed. Thus
j
=
F( y)
j
.
Chapter 10
Perturbation theory
10.1 The implicit function theorem: scalar case
Let f(u, ) be a function of u and . Let f
u
(u, ) be the partial derivative of F
as a function of u. The assumption for the implicit function theorem is that
f(u
0
, 0) = 0 and the partial derivative f
u
(u
0
, 0) ,= 0. There are some other
technical assumptions. The conclusion is that the equation
f(u, ) = 0 (10.1)
denes u as a function of for suciently near 0, in such a way that = 0 is
mapped into u
0
.
The intuition behind the theorem is that one can expand
f(u, ) f
u
(u
0
, 0)(u u
0
) +f
(u
0
, 0). (10.2)
Then we can solve
u u
0
f
u
(u
0
, 0)
1
f
(u
0
, u). (10.3)
For small this is a very good approximation. In fact, this says that at the
point = 0 we have
du
d
= f
u
(u
0
, 0)
1
f
(u
0
, 0). (10.4)
This result is called rst order perturbation theory.
For practical calculations one wants to compute higher derivatives u
(n)
as a
function of . These are used in the Taylor expansion
u =
n=0
u
(n)
n!

n
. (10.5)
In rst order we have
f
u
u
(1)
+f
= 0 (10.6)
133
134 CHAPTER 10. PERTURBATION THEORY
which gives the rst order result
u
(1)
= f
1
u
f
. (10.7)
Usually we are interested in evaluating this expression at u = u
0
and = 0.
The second order result is obtained by dierentiating the rst order equation.
This gives
f
u
u
(2)
+f
uu
(u
(1)
)
2
+ 2f
u
u
(1)
+f
= 0. (10.8)
Thus the result of second order perturbation theory is that
u
(2)
= f
1
u
[f
uu
(u
(1)
)
2
+ 2f
u
u
(1)
+f
]. (10.9)
Again we are interested in evaluating this expression at u = u
0
and = 0.
The general pattern is this. The nth derivative is given by
f
u
u
(n)
+r
n
= 0 (10.10)
Dierentiate again. This gives
f
u
u
(n+1)
+ [f
uu
u
(1)
+f
u
]u
(n)
+ [r
nu
u
(1)
+r
n
] = 0. (10.11)
Thus at every stage one only has to invert the rst derivative f
u
.
Example: Here is a simple example. Fix a. We want to solve u = a +g(u)
to nd u as a function of . The rst derivative is u
= g(u) + g
(u)u
. The
second derivative is u
= 2g
(u)u
+g
(u)u
+g
(u)u
2
. The third derivative

is u
= 3g
(u)u
2
+3g
(u)u
+ times other junk. Evaluate these at = 0, that

is, at u = a. This gives u
= g(a), u
= 2g(a)g
(a), and u
= 3g(a)
2
g
(a) +
6g(a)g
(a)
2
. It is dicult to see the pattern. However in the next section we
will see that there is one, at least in this case.
Example: Another popular way of solving the problem is to substitute u =
a +u
+
1
2
u
2
+
1
6
u
3
+ in the equation. This gives
a+u
+
1
2
u
2
+
1
6
u
3
+ = a+[g(a)+g
(a)(u
+
1
2
u
2
+ )+
1
2
2
g
(a)(u
+ )
2
+ ].
(10.12)
Equating powers of gives the same equations, and these are solved recursively
the same way.
10.2 Problems
There is at least one case when one can evaluate Taylor coecients to all orders.
Say that x is a xed parameter and one wants to solve
u = x +tg(u) (10.13)
for u as a function of t. When t = 0 the solution is u = x. The Lagrange
expansion says that for arbitrary n 1 we have
d
n
u
dt
n
[
t=0
=
d
n1
dx
n1
g(x)
n
. (10.14)
10.2. PROBLEMS 135
So the expansion to all orders is
u = x+tg(x) +
t
2
2
d
dx
g(x)
2
+
t
3
6
d
2
dx
2
g(x)
3
+ +
t
n
n!
d
n1
dx
n1
g(x)
n
+ . (10.15)
1. Let w = g(u). Show that the equation becomes
w = g(x +wt). (10.16)
2. Show that the function w is the solution of the partial dierential equation
w
t
= w
w
x
(10.17)
with initial condition w = g(x) at t = 0.
3. Prove the conservation laws
w
n
t
=
n
n + 1
w
n+1
x
. (10.18)
4. Prove that for n 1 we have
n1
w
t
n1
=
1
n
n1
x
n1
w
n
. (10.19)
5. Prove that at t = 0
w = g(x)+
t
2
d
dx
g(x)
2
+
t
2
6
d
2
dx
2
g(x)
3
+ +
t
n1
n!
d
n1
dx
n1
g(x)
n
+ . (10.20)
6. Use u = x +tw to prove the Lagrange expansion.
Note: The partial dierential equation has a physical interpretation. If
v = w, then the equation is
v
t
+v
v
x
= 0 (10.21)
with initial condition v = h(x) at time zero. This is the equation for the velocity
of a gas of particles moving freely in one dimension. The motion of a particle
moving with the gas is given by
dx
dt
= v. (10.22)
The acceleration of a particle is
dv
dt
=
v
t
+
dx
dt
v
x
= 0 (10.23)
That is why it is considered free particle motion. Since v is constant along
particle paths x = x
0
+ vt, we have that v at arbitrary x, t is given in terms of
the initial v = h(x
0
) at time zero by v = h(x
0
) = h(x vt). So the solution of
the equation is
v = h(x vt). (10.24)
10.3 The implicit function theorem: systems
We can take F(u, ) = 0 to be a system of n equations for n variables u and an
extra variable . Or more generally we can take F(u, ) to be a function of u
and , where u ranges over a region U in some Banach space E. Then F(u, )
takes values in a Banach space E
. The parameter is still real.

We shall often abbreviate the function by F. Let F
u
be the partial derivative
of F as a function of u. In the case of a system this is the n by n matrix of
partial derivatives of the components of F with respect to the components of u.
More generally, it is a function on U with values that are linear transformations
from E to E
.
The assumption for the implicit function theorem is that F(u
0
, 0) = 0 and
the partial derivative F
u
(u
0
, 0) has a bounded inverse. There are some other
technical assumptions. The conclusion is that the equation
F(u, ) = 0 (10.25)
denes u as a function of for suciently near 0, in such a way that = 0 is
mapped into u
0
.
For practical calculations one wants to compute higher derivatives u
(n)
as a
function of . These are used in the Taylor expansion
u =
n=0
u
(n)
n!

n
. (10.26)
In rst order we have
F
u
u
(1)
+F
= 0 (10.27)
which gives the rst order result
u
(1)
= F
1
u
F
. (10.28)
Note that F
has values in E
. On the other hand, F

1
u
has values that are
linear transformation from E
to E. So the right hand side has values in E, as

it must. Usually we are interested in evaluating this expression at u = u
0
and
= 0.
The second order result is obtained by dierentiating the rst order equation.
This gives
F
u
u
(2)
+F
uu
(u
(1)
, u
1)
) + 2F
u
u
(1)
+F
= 0. (10.29)
Thus the result of second order perturbation theory is that
u
(2)
= F
1
u
[F
uu
(u
(1)
, u
1)
) + 2F
u
u
(1)
+F
]. (10.30)
Note that F
uu
has values that are bilinear functions from E E to E
. On
the other hand, F
u
has values that are linear functions from E to E
. Again
the right hand side has values in E. Again we are interested in evaluating this
expression at u = u
0
and = 0.
10.4. NONLINEAR DIFFERENTIAL EQUATIONS 137
The general pattern is this. The nth derivative is given by
F
u
u
(n)
+R
n
= 0 (10.31)
Dierentiate again. This gives
F
u
u
(n+1)
+ [F
uu
u
(1)
+F
u
]u
(n)
+ [R
nu
u
(1)
+R
n
] = 0. (10.32)
Thus at every stage one only has to invert the rst derivative F
u
.
10.4 Nonlinear dierential equations
Let us look at the simplest non-linear dierential equation
Lu = g(u). (10.33)
Here L is the restriction of a linear operator to satisfy an inhomogeneous bound-
ary condition. For example, we could have L = /t+1 acting on functions with
u(0) = 1. When = 0 there is a solution u = u
0
. In the example u
0
(t) = e
t
.
The derivative of L is L
0
, where L
0
is a linear operator with zero boundary
conditions. For instance, we could have L
0
= /t+1 acting on functions h with
h(0) = 0. Thus the rst order perturbation equation is L
0
u
= g(u) +g
(u)u
.
The second order equation is L
0
u
= 2g
(u)u
+ terms with . One can continue

in this way. Thus the hierarchy of equations begins Lu
0
= 0, L
0
u
= g(u
0
),
L
0
u
= 2g
(u
0
)u
.
It is possible to think of this in another way. Write u = u
0
+ w. Then the
equation Lu = g(u) becomes L
0
w = g(u), where L
0
is the operator dened
with zero boundary conditions. Suppose that L
0
has an inverse. Then
u = u
0
+L
1
0
g(u). (10.34)
This is very close in form to the scalar equation examined in the rst two
sections.
This equation may be solved by the usual procedure. The rst derivative
is u
= L
1
0
g(u) + L
1
0
g
(u)u
. The second derivative is u
= 2L
1
0
g
(u)u
+
L
1
0
g
(u)u
+ L
1
0
[g
(u)u
2
]. The third derivative is u
= 3L
1
0
[g
(u)u
2
] +
3L
1
0
[g
(u)u
] + times other junk. Evaluate these at = 0, that is, at u = u

0
.
This gives
u
= L
1
0
g(u
0
), (10.35)
and
u
= 2L
1
0
[g
(u
0
)L
1
0
g(u
0
)], (10.36)
and
u
= 3L
1
0
[g
(u
0
)(L
1
0
(g(u
0
))
2
] + 6L
1
0
[g
(u
0
)L
1
0
(g
(u
0
)L
1
0
g(u
0
))]. (10.37)
Notice how the linear operators occur in various places, making it impossible to
do the simplication that we found in the scalar case.
10.5 A singular perturbation example
In quantum mechanics one often wants to solve a problem of the form
h
2
2m
u
+v(x)u = u. (10.38)
Here v(x) is a given real function. In many contexts plays the role of an
eigenvalue parameter. The constant m > 0 is xed. The constant h > 0 is the
perturbation parameter.
The characteristic feature of this problem is that the zero order perturbation
is not apparent. Thus setting h = 0 it is not clear that there will be a solution
at all.
The resolution of this problem is to write the equation in new variables. For
simplicity think of the case when > v(x). Thus
u = Ae
i
S
h
. (10.39)
The equation becomes
A
_
1
2m
(S
)
2
+v(x)
_
=
h
2
2m
A
(10.40)
together with
2A
+AS
= 0. (10.41)
Now we can recognize the h = 0 limit as
1
2m
(S
)
2
+v(x) = . (10.42)
The other equation may be written
(A
2
S
= 0. (10.43)
These equations have a physical interpretation. The function S
is a velocity
that depends on position. The rst equation is the classical equation for con-
servation of energy. The function A
2
is a density. The second equation is the
conservation law that says that the current A
2
S
is constant in space. Of course

these equations are only valid in the zero order approximation, which in this
case is called the WKB approximation.
This approximation is relevant to a heuristic understanding of Schrodinger
operators of the form
H =
h
2
2m
d
2
dx
2
+v(x). (10.44)
The Hilbert space is the space of L
2
functions on the line. If v(x) is bounded
below, then one expects this to be a well-dened self-adjoint operator. In fact,
for < 0 suciently negative, one expects (H I)
1
to be a well-dened
bounded self-adjoint operator. The reason is that the solutions of the dierential
10.6. EIGENVALUES AND EIGENVECTORS 139
equation either decay rapidly at innity or grow rapidly at innity. The Greens
function is dened by the solutions that decay rapidly at innity. There is a
solution of Hu = u that is in L
2
at and by another solution of the same
equation that is in L
2
at +. The two solutions are matched by a jump
condition to form the kernel of the integral operator.
When v(x) is very badly unbounded below, then something quite dierent
happens. Let us look at the example of v(x) x
2n
with < 0. According to
the WKB approximation, we have S
(x) [x[
n
. It follows that A
2
[x[
n
.
Thus it is plausible that if n > 1 then every solution is in L
2
. These solutions
are highly oscillatory and decrease at innity at a moderate rate.
If this is indeed the case, then the operator H is not self-adjoint. The
problem is that when one wants to dene a Greens function, there are two
linearly independent solutions, and they are both in L
2
at . So there is no
boundary condition at innity that determines the Greens function in a natural
way.
10.6 Eigenvalues and eigenvectors
We have an operator H
0
+ V . We are interested in an isolated eigenvalue
of multiplicity 1. The corresponding eigenvector is p. It is only determined up
to a constant multiple. The adjoint operator has eigenvalue

with eigenvalue
q. It is convenient to choose p and q so that the inner product (q, p) = 1. The
associated projection onto the eigenspace is P satisfying P
2
= P. It is given by
Pu = (q, u)p. Thus p is an eigenvector with
(H
0
+V )p = 0. (10.45)
Furthermore, (q, (H
0
+ V )u) = 0 for all u. We also have the reduced
resolvent S with SP = PS = 0 and S(H
0
+V I) = 1 P.
If we indicate the dependence on explicitly, then the equation we want to
solve is
(H
0
+V ())p() = 0. (10.46)
We would like to nd the perturbation expansion in terms of the and p and q
and S associated with = 0.
The following discussion is meant to establish two formulas, one for the
eigenvalue and one for the eigenvector (not normalized). The rst is the formula
for the eigenvalue to second order
() = + (q, V p) (q, V SV p)
2
+ . (10.47)
The other is the formula for the (unnormalized) eigenvector to second order
p() = p SV p + [SV SV p S
2
V p(q, V p)]
2
+ . (10.48)
These formulas are so important that they should be memorized, at least the
one for the eigenvalues.
We now proceed to the derivations. Dierentiate the eigenvector equation
with respect to . This gives
(H
0
+V ) p

p +V p = 0. (10.49)
Now take the inner product with q on the left. This gives
= (q, V p). (10.50)

This gives the rst order change of the eigenvalue. That is,
() = + (q, V p) . (10.51)
Next apply I P on the left. This gives
(I P)(H
0
+V ) p + (I P)V p = 0, (10.52)
or
(I P) p = SV p. (10.53)
Thus
p = SV p +c
1
p. (10.54)
Notice that the derivative of the eigenvector is not uniquely determined. It
ordinarily harmless to take c
1
= 0. This gives the rst order change of the
eigenvector
p() = p + [SV p +c
1
p]. (10.55)
Now dierentiate again. This gives
(H
0
+V ) p +

p + 2V p 2
p = 0. (10.56)
Take the inner product with q on the left. This gives
+ 2(q, V p) 2
(q, p), (10.57)

or
1
2
= (q, V SV p). (10.58)

This gives the second order change of the eigenvalue. That is,
() = + (q, V p) (q, V SV p)
2
+ . (10.59)
Now apply 1 P on the left. This gives
(1 P)(H
0
+V ) p + 2(1 P)V p 2
(1 P) p = 0. (10.60)
Thus
1
2
(1 P) p = SV SV p
1
2
c
1
SV p S
2
V p(q, V p). (10.61)
10.7. THE SELF-ADJOINT CASE 141
Hence
1
2
p = SV SV p S
2
V p(q, V p)
1
2
c
1
SV p +
1
2
c
2
p. (10.62)
This gives the second order change of the eigenvector
p() = p +[SV p +c
1
p] +[SV SV p S
2
V p(q, V p)
1
2
c
1
SV p +
1
2
c
2
p]
2
+ .
(10.63)
Now let us change notation. We will think of a basis p
n
for the Hilbert space
and a dual basis q
m
, with (q
m
, p
n
) =
mn
. We write q
m
for the functional that
takes u to (q
m
, u). Then the various operators have special forms. We take the
p
n
to be the eigenvectors of H
0
. Thus H
0
p
n
=
n
p
n
. The operator P
n
= p
n
q
n
is the spectral projection associated with the nth eigenvalue
n
of H
0
. The
operator S
n
=

j=n
p
j
q
j
/(
k

n
).
Then the second order expression for the eigenvalue can be written as
n
() =
n
+ (q
n
, V p
n
) (q
n
, V S
n
V p
n
)
2
+ . (10.64)
Even more explicitly, this is
n
() =
n
+ (q
n
, V p
n
)
j=n
(q
n
, V p
j
)
1
j

n
(q
j
, V p
n
)
2
+ . (10.65)
The second order expression for the eigenvector p
n
() is (taking c
1
= 0)
p
n
() = p
n
S
n
V p
n
+[S
n
V S
n
V p
n
S
2
n
V p
n
(q
n
, V p
n
)+
1
2
c
2
p
n
]
2
+ . (10.66)
If H
0
has discrete spectrum, then (leaving out the last term) this may be written
p
n
() = p
n
j=n
1
j

n
(q
j
, V p
n
)p
j
(10.67)
+
j=n
1
j

n
[
k=n
(q
j
, V p
k
)
1
k

n
(q
k
, V p
n
)
1
j

n
(q
j
, V p
n
)(q
n
, V p
n
)]p
j
2
+ .
10.7 The self-adjoint case
In the self-adjoint case q = p and it is natural to normalize so that (p(), p()) = 1
to second order. This gives c
1
= 0 and
c
2
= (p, V S
2
V p). (10.68)
So in this case
() = + (p, V p) (p, V SV p)
2
+ . (10.69)
and
p() = p SV p + [SV SV p S
2
V p(p, V p)
1
2
(p, V S
2
V p)p]
2
+ . (10.70)
If =
n
and p = p
n
, we can write the coecient c
2
even more explicitly as
c
2
=
j=n
(p
n
, V p
j
)
1
(
j

n
)
2
(p
j
, V p
n
). (10.71)
Example: Take
H
0
=
_
1 0
0 1
_
(10.72)
and
V =
_
0 1
1 0
_
. (10.73)
The eigenvalues are () =
1 +
2
. However it is interesting to compare this
exact result with the result of second order perturbation theory.
Let us look at the perturbation of the eigenvalue 1 of H
0
. For this eigen-
value the spectral projection is
P =
_
0 0
0 1
_
(10.74)
and the eigenvector p is the second column of P. The reduced resolvent is
S =
_
1
1(1)
0
0 0
_
=
_
1
2
0
0 0
_
. (10.75)
So the coecient in second order perturbation theory is
(p, V SV p) =
1
2
. (10.76)
Thus
() = 1
1
2
2
+ . (10.77)
10.8 The anharmonic oscillator
In this section we let
H
0
=
1
2
_
d
2
dx
2
+x
2
1
_
. (10.78)
The perturbation V is multiplication by
v(x) = x
4
. (10.79)
Let
A =
1
2
_
x +
d
dx
_
(10.80)
and
A
=
1
2
_
x
d
dx
_
. (10.81)
10.8. THE ANHARMONIC OSCILLATOR 143
Then AA
= A
A+I and H
0
= A
A. Furthermore, the solution of Au

0
= 0 is
u
0
(x) =
1
4
x
2
2
. (10.82)
Dene for each n = 0, 1, 2, 3, . . . the vector
u
n
=
1
n!
A
n
u
0
. (10.83)
We have A
u
m
=

m+ 1u
m+1
and Au
m
=

mu
m1
. Then the u
n
form an
orthonormal basis consisting of eigenvectors of H
0
. Furthermore,
H
0
u
n
= nu
n
. (10.84)
Say that we are interested in a particular n. The corresponding reduced
resolvent S
n
is given by
S
n
u
m
=
1
mn
u
m
(10.85)
for m ,= n, and by S
n
u
n
= 0.
The operator multiplication by x may be expressed in terms of A and A
by
x =
1
2
(A+A
). (10.86)
This allows us to calculate the matrix elements of x. These are given by
(u
m+1
, xu
m
) =
1
2
(u
m+1
, A
u
m
) =
1
m+ 1 (10.87)
and
(u
m1
, xu
m
) =
1
2
(u
m1
, Au
m
) =
1
m. (10.88)
So each transition is to a neighboring natural number.
As a warmup, let us compute the matrix element (u
n
, x
2
u
n
). This is the sum
of two possible transition paths, one up-down and one down-up. The result is
(n + 1)/2 + (n/2) = (2n + 1)/2.
To compute the rst order perturbation coecient (u
n
, x
4
u
n
). we need to
look at all transitions that go from n to n in 4 steps. After two steps one is
either at u
n2
, at u
n
, or at u
n+2
. So
(u
n
, x
4
u
n
) = (u
n
, x
2
u
n2
)
2
+ (u
n
, x
2
u
n
)
2
+ (u
n
, x
2
u
n+2
)
2
. (10.89)
This is
(u
n
, x
4
u
n
) =
n(n 1)
4
+ (
2n + 1
2
)
2
+
(n + 1)(n + 2)
4
=
6n
2
+ 6n + 3
4
. (10.90)
Higher order perturbations may be computed with more labor. It turns out
that the perturbation series in does not converge. Some intuition for this is
provided by the observation that H
0
+V is not a self-adjoint operator for < 0.
So for < 0 the eigenvalue problem is meaningless. On the other hand, if the
Taylor series had a positive radius of convergence, then it would converge also
for some values of < 0.

Appliid Math Text

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Appliid Math Text

Diunggah oleh

Hak Cipta:

Format Tersedia

Methods of Applied Mathematics

Ax. Then the equation says that one

. In the real case this is the same

. Then there is a unitary

= A) is normal. A skew-adjoint matrix (A

of a vector space V consists of all linear transformations

, then the value of on u in V is

is via its contour lines. These are

are called covariant vectors.

in the dual space of V dened by

= gu is a form with contour lines perpendicular to

, z)c = g(z, cu). Thus justies the denition u

such that g(c

, v) = c, v). This justies the denition

t), we get the

g denotes the square root of the determinant of the metric form

< /2. (2.21)

norm by coecients that vanish at innity. Therefore the coecients

exists (in the sense that f is an integral

(x). So the hypothesis of the theorem is a condition related to dierentiability

parameterized by k. The goal is to

) be the space of all (Borel measurable) complex functions such that

) is a Hilbert space with inner product

). So this theory is simple and powerful.

, the space of essentially bounded functions. An example of a

f(x y)g(y) dy. (3.8)

(x) = f(x). Then the Fourier transform of f

), the space of bounded

(k) = exp([k[). (3.11)

is an approximate delta function. The result follows by taking 0.

in the same way, so that if h is in

h, f) = (h, Ff). Take h = Fg. Then

Fg, f) = (Fg, Ff) = (g, f). That is F

Fg = g. Similarly, one may show

u = u. These equations show that F is unitary and that F

). The inverse Fourier transform F

). For the other spaces the situation is more complex.

). That is, the Fourier transform of an inte-

) function, and therefore is in L

be an approximate function. Then (

exists (in the sense that f is an integral

(k) by the usual formula. Check that

and its conjugate are not in L

(R) norms of f, and compare them.

) norms of the Fourier transform, and

) norms for this problem.

) and such that

(z) be a small circle about z. Let R

be the region R with the

(z) the absolute value [f()f(z)[ . Then

be R with the singularities omitted. Consider small circles

z. The most natural denition of this

f(x)(x) dx. (5.1)

(x a)(x) dx. (5.4)

(x) = PV 1/x. Here the principal value PV 1/x is given by

. This is dened in the obvious way by

(x) = (x/)/. Then the functions

converge to the distribution as

has integral one. Consider a

0 be a positive function with integral 1.

converge to the distribution as tends to

(x) dx. Then

(a) 0, and for each a > 0 we have 1 H

(a) H(a) as 0. Since the